Mathematical Physics A Modern Intro to its Foundations S

Sadri Hassani Mathematical Physics A Modem Introduction to Its Foundations With 152 Figures , Springer ODTlJ KU1"UPHA...

1 downloads 60 Views 34MB Size
Sadri Hassani

Mathematical Physics A Modem Introduction to Its Foundations With 152 Figures

,

Springer ODTlJ KU1"UPHANESt M. E. T. U. liBRARY

Sadri Hassani Department of Physics

Illinois State University Normal, IL 61790 USA [email protected]

QC20 H394

QC20 14394

METULIBRARY

c,2.

.. 2 ~themltlcoJ

To my wife Sarah

physicS: I modem

ml l l l l l l l l l l002m7S69 l l l l ~1 1 1 1 1 1 1 1 1 1 1 1 1 I I

and to my children

Dane Arash and Daisy Rita

336417 Libraryof CongressCataloging-in-Publication Data Hassani, Sadri. Mathematical physics: a modem introductionits foundations / Sadri Hassani. p. em. Includesbibliographical referencesand index. ISBN 0-387-98579-4 (alk. paper) 1. Mathematical physics. I. Title. QC20.H394 1998 530.15--<1c21 98-24738 Printed on acid-freepaper. © 1999Springer-Verlag New York,Inc. All rights reserved.This work may not be translatedor copied in whole or in part without the written permission of the publisher (Springer-Verlag New York, Inc., 175 Fifth Avenue, New York, NY 10010, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use-in connection with any form of informationstorage and retrieval, electronicadaptation, computer scttware, or by similaror dissimilarmethodology now knownor hereafterdevelopedis forbidden. The use of general descriptivenames, trade names, trademarks,etc., in this publication, even if the formerare not especiallyidentified, is not to be taken as a sign that such names, as understoodby the Trade Marks and Merchandise Marks Act, may accordinglybe used freely by anyone.

Productionmanagedby Karina Mikhli;manufacturing supervisedby ThomasKing. Photocomposed copy preparedfrom the author's TeX files. Printed and bound by HamiltonPrinting Co., Rensselaer, NY. Printed in the United Statesof America. 9 8 7 6 5 4 3 (Correctedthird printing, 2002) ISBN 0-387-98579-4

SPIN 10854281

Springer-Verlag New York Berlin: Heidelberg A member. ofBertelsmannSpringer Science+Business Media GmbH

Preface

"Ich kann es nun einmal nicht lassen, in diesem Drama von Mathematik und Physik---
Hermann Weyl

It is said that mathematics is the language of Nature. If so, then physics is its poetry. Nature started to whisper into our ears when Egyptians and Babylonians were compelled to invent and use mathematics in their day-to-day activities. The faint geomettic and arithmetical pidgin of over four thousand years ago, snitable for rudimentary conversations with nature as applied to simple landscaping, has turned into a sophisticated language in which the heart of matter is articulated. The interplay between mathematics and physics needs no emphasis. What may need to be emphasized is that mathematics is not merely a tool with which the presentation of physics is facilitated, but the only medium in which physics can survive. Just as language is the means by which humans can express their thoughts and without which they lose their unique identity, mathematics is the only language through which physics can express itself and without which it loses its identity. And just as language is perfected due to its constant usage, mathematics develops in the most dramatic way because of its usage in physics. The quotation by Weyl above, an approximation to whose translation is "In this drama of mathematics and physics-which fertilize each other in the dark, but which prefer to deny and misconstrue each other face to face-I cannot, however, resist playing the role of a messenger, albeit, as I have abundantly learned, often an unwelcome one:'

vi

PREFACE

is a perfect description of the natnral intimacy between what mathematicians and physicists do, and the nnnatnral estrangement between the two camps. Some ofthe most beantifnl mathematics has been motivated by physics (differential eqnations by Newtonian mechanics, differential geometry by general relativity, and operator theory by qnantnmmechanics), and some of the most fundamental physics has been expressed in the most beantiful poetry of mathematics (mechanics in symplectic geometry, and fundamental forces in Lie group theory). I do uot want to give the impression that mathematics and physics cannot develop independently. On the contrary, it is precisely the independence of each discipline that reinforces not only itself, but the other discipline as well-just as the stndy of the grammar of a language improves its usage and vice versa. However, the most effective means by which the two camps can accomplish great success is throngh an inteuse dialogue. Fortnnately, with the advent of gauge and string theories ofparticle physics, such a dialogue has beenreestablished between physics and mathematics after a relatively long lull.

Level and Philosophy of Presentation This is a book for physics stndeuts interested in the mathematics they use. It is also a book fur mathematics stndeuts who wish to see some of the abstract ideas with which they are fantiliar come alive in an applied setting. The level of preseutation is that of an advanced undergraduate or beginning graduate course (or sequence of courses) traditionally called "Mathematical Methods of Physics" or some variation thereof. Unlike most existing mathematical physics books intended for the same audience, which are usually lexicographic collections of facts about the diagonalization of matrices, tensor analysis, Legendre polynomials, contour integration, etc., with little emphasis on formal and systematic development of topics, this book attempts to strike a balance between formalism and application, between the abstract and the concrete. I have tried to include as mnch of the essential formalism as is necessary to render the book optimally coherent and self-contained. This entails stating and proving a large nnmber of theorems, propositions, lemmas, and corollaries. The benefit of such an approach is that the stndent will recognize clearly both the power and the limitation of a mathematical idea used in physics. There is a tendency on the part of the uovice to universalize the mathematical methods and ideas eucountered in physics courses because the limitations of these methods and ideas are not clearly pointed out. There is a great deal of freedom in the topics and the level of presentation that instructors can choose from this book. My experience has showu that Parts I, TI, Ill, Chapter 12, selected sections of Chapter 13, and selected sections or examples of Chapter 19 (or a large snbset of all this) will be a reasonable course content for advanced undergraduates. Ifone adds Chapters 14 and 20, as well as selected topics from Chapters 21 and 22, one can design a course snitable for first-year graduate

PREFACE

vii

students. By judicious choice of topics from Parts VII and VIII, the instructor can bring the content of the course to a more modern setting. Depending on the sophistication of the students, this can be done either in the first year or the second year of graduate school.

Features To betler understand theorems, propositions, and so forth, students need to see them in action. There are over 350 worked-out examples and over 850 problems (many with detailed hints) in this book, providing a vast arena in which students can watch the formalism unfold. The philosophy underlying this abundance can be summarized as ''An example is worth a thousand words of explanation." Thus, whenever a statement is intrinsically vague or hard to grasp, worked-out examples and/or problems with hints are provided to clarify it. The inclusion of such a large number of examples is the means by which the balance between formalism and application has been achieved. However, although applications are essential in understanding mathematical physics, they are only one side of the coin. The theorems, propositions, lemmas, and corollaries, being highly condensed versions of knowledge, are equally important. A conspicuous feature of the book, which is not emphasized in other comparable books, is the attempt to exhibit-as much as.it is useful and applicable-« interrelationships among various topics covered. Thus, the underlying theme of a vector space (which, in my opinion, is the most primitive concept at this level of presentation) recurs throughout the book and alerts the reader to the connection between various seemingly unrelated topics. Another useful feature is the presentation of the historical setting in which men and women of mathematics and physics worked. I have gone against the trend of the "ahistoricism" of mathematicians and physicists by summarizing the life stories of the people behind the ideas. Many a time, the anecdotes and the historical circumstances in which a mathematical or physical idea takes form can go a long way toward helping us understand and appreciate the idea, especially if the interaction among-and the contributions of-all those having a share in the creation of the idea is pointed out, and the historical continuity of the development of the idea is emphasized. To facilitate reference to them, all mathematical statements (definitions, theorems, propositions, lemmas, corollaries, and examples) have been nnmbered consecutively within each section and are preceded by the section number. For example, 4.2.9 Definition indicates the ninth mathematical statement (which happens to be a definition) in Section 4.2. The end of a proof is marked by an empty square D, and that of an example by a filled square III, placed at the right margin of each. Finally, a comprehensive index, a large number of marginal notes, and many explanatory underbraced and overbraced comments in equations facilitate the use

viii

PREFACE

and comprehension of the book. In this respect, the book is also nsefnl as a reference.

Organization and Topical Coverage Aside from Chapter 0, which is a collection of pnrely mathematical concepts, the book is divided into eight parts. Part I, consisting of the first fonr chapters, is devoted to a thorough study of finite-dimensional vector spaces and linear operators defined on them. As the unifying theme of the book, vector spaces demand careful analysis, and Part I provides this in the more accessible setting of finite dimension in a language thatis conveniently generalized to the more relevant infinite dimensions,' the subject of the next part. Following a brief discussion of the technical difficulties associated with infinity, Part IT is devoted to the two main infinite-dimensional vector spaces of mathematical physics: the classical orthogonal polynomials, and Foutier series and transform. Complex variables appear in Part ill. Chapter 9 deals with basic properties of complex functions, complex series, and their convergence. Chapter 10 discusses the calculus of residues and its application to the evaluation of definite integrals. Chapter II deals with more advanced topics such as multivalued functions, analytic continuation, and the method of steepest descent. Part IV treats mainly ordinary differential equations. Chapter 12 shows how ordinary differential equations of second order arise in physical problems, and Chapter 13 consists of a formal discussion of these differential equations as well as methods of solving them numerically. Chapter 14 brings in the power of complex analysis to a treatment of the hypergeometric differential equation. The last chapter of this part deals with the solution of differential equations using integral transforms. Part V starts with a formal chapter on the theory of operator and their spectral decomposition in Chapter 16. Chapter 17 focuses on a specific type of operator, namely the integral operators and their corresponding integral equations. The formalism and applications of Stnrm-Liouville theory appear in Chapters 18 and 19, respectively. The entire Part VI is devoted to a discussion of Green's functions. Chapter 20 introduces these functions for ordinary differential equations, while Chapters 21 and 22 discuss the Green's functions in an m-dimensional Euclidean space. Some of the derivations in these last two chapters are new and, as far as I know, unavailable anywhere else. Parts VII and vrn contain a thorough discussion of Lie groups and their applications. The concept of group is introduced in Chapter 23. The theory of group representation, with an eye on its application in quantom mechanics, is discussed in the next chapter. Chapters 25 and 26 concentrate on tensor algebra and ten-, sor analysis on manifolds. In Part vrn, the concepts of group and manifold are

PREFACE

ix

brought together in the coutext of Lie groups. Chapter 27 discusses Lie groups and their algebras as well as their represeutations, with special emphasis on their application in physics. Chapter 28 is on differential geometry including a brief introduction to general relativity. Lie's original motivation for constructing the groups that bear his name is discussed in Chapter 29 in the context of a systematic treatment of differential equations using their symmetry groups. The book ends in a chapter that blends many of the ideas developed throughout the previous parts in order to treat variational problems and their symmetries. It also provides a most fitting example of the claim made at the beginning of this preface and one of the most beautiful results of mathematical physics: Noether's theorem ou the relation between symmetries and conservation laws.

Acknowledgments It gives me great pleasure to thank all those who contributed to the making of this book. George Rutherford was kind enough to voluuteer for the difficult task of condensing hundreds of pages of biography into tens of extremely informative pages. Without his help this unique and valuable feature of the book would have been next to impossible to achieve. I thank him wholeheartedly. Rainer Grobe and Qichang Su helped me with my rusty computational skills. (R. G. also helped me with my rusty German!) Many colleagues outside my department gave valuable comments and stimulating words of encouragement on the earlier version of the book. I would like to record my appreciation to Neil Rasband for reading part of the manuscript and commenting on it. Special thanks go to Tom von Foerster, senior editor of physics and mathematics at Springer-Verlag, not ouly for his patience and support, but also for the,extreme care he took in reading the entire manuscript and giving me invaluable advice as a result. Needless to say, the ultimate responsibility for the content of the book rests on me. Last but not least, I thank my wife, Sarah, my son, Dane, and my daughter, Daisy, for the time taken away from them while I was writing the book, and for their support during the long and arduous writing process. Many excellent textbooks, too numerous to cite individually here, have influenced the writing of this book. The following, however, are noteworthy for both their excellence and the amount of their influence: Birkhoff, G., and G.-C. Rota, Ordinary Differential Equations, 3rd ed., New York, Wiley, 1978. Bishop, R., and S. Goldberg, Tensor Analysis on Manifolds, New York, Dover, 1980. Dennery, P., and A. Krzywicki, Mathematics for Physicists, New York, Harper & Row, 1967. Halmos, P., Finite-Dimensional Vector Spaces, 2nd ed., Princeton, Van Nostrand, 1958.

x

PREFACE

Hamennesh, M. Group Theory and its Application to Physical Problems, Dover, New York, 1989. Olver, P.Application ofLie Groups to Differential Equations, New York, SpringerVerlag, 1986. Unless otherwise indicated, all biographical sketches have been taken from the following three sources: Gillispie, C., ed., Dictionary ofScientific Biography, Charles Scribner's, New York, 1970. Simmons, G. Calculus Gems, New York, McGraw-Hill, 1992. History of Mathematics archive at www-groups.dcs.st-and.ac.uk:80. I wonld greatly appreciate any comments and suggestions for improvements. Although extreme care was taken to correct all the misprints, the mere volume of the book makes it very likely that I have missed some (perhaps many) of them. I shall be most grateful to those readers kind enough to bring to my attention any remaining mistakes, typographical or otherwise. Please feel free to contact me. Sadri Hassani Campus Box 4560 Department of Physics Illinois State University Normal, IL 61790-4560, USA e-mail: [email protected]

It is my pleasure to thank all those readers who pointed out typographical mistakes and suggested a few clarifying changes. With the exception of a couple that required substantial revisiou, I have incorporated all the corrections and suggestions in this second printing.

Note to the Reader

Mathematics and physics are like the game of chess (or, for that matter, like any gamej-i-you willleam only by ''playing'' them. No amount of reading about the game will make you a master. In this book you will find a large number ofexamples and problems. Go through as many examples as possible, and try to reproduce them. Pay particular attention to sentences like "The reader may check . .. "or "It is straightforward to show . .. "These are red flags warning you that for a good understanding of the material at hand, yon need to provide the missing steps. The problems often fill in missing steps as well; and in this respect they are essential for a thorough understanding of the book. Do not get discouraged if you cannot get to the solution of a problem at your first attempt. If you start from the beginning and think about each problem hard enough, you will get to the solution, .and you will see that the subsequent problems will not be as difficult. The extensive index makes the specific topics about which you may be interested to leam easily accessible. Often the marginal notes will help you easily locate the index entry you are after. I have included a large collection of biographical sketches of mathematical physicists ofthe past. These are truly inspiring stories, and I encourage you to read them. They let you see that even under excruciating circumstances, the human mind can work miracles. You will discover how these remarkable individuals overcame the political, social, and economic conditions of their time to let us get a faint glimpse of the truth. They are our true heroes.

1

1

1

1

1

1

1

1

1

1

1

1

1

1

1

1

Contents

Preface Note to the Reader List of Symbols

o

Mathematical Preliminaries 0.1 Sets . 0.2 Maps . 0.3 Metric Spaces . 0.4 Cardinality........ 0.5 Mathematical Induction 0.6 Problems..........................

v

xi xxi

1 1 4 7 10

12 14

I Finite-Dimensional Vector Spaces

17

1 Vectors and Transformations 1.1 Vector Spaces . . . . . 1.2 Inner Product . 1.3 Linear Transformations . 1.4 Algebras. 1.5 Problems..........

19

2 Operator Algebra 2.1 Algebra «s: (V)

19

23 32 41 44

49 49

xiv

CONTENTS

2.2 2.3 2.4 2.5 2.6 2.7

Derivatives of Functions of Operators . Conjugation of Operators . . . . Hermitian and Unitary Operators Projection Operators . . . . . . . Operators in Numerical Analysis Problems.............

3 Matrices: Operator Representations 3.1 Matrices.......... 3.2 Operations on Matrices .. . . . . 3.3 OrthonormalBases . 3.4 Change of Basis and Similarity Transformation . 3.5 The Determinant . . .. 3.6 The Trace 3.7 Problems...... 4

Spectral Decomposition 4.1 Direct Sums . . . . . . . . . . 4.2 InvariantSubspaces . . . . . . 4.3 Eigeuvaluesand Eigenvectors . 4.4 Spectral Decomposition 4.5 Functions of Operators 4.6 Polar Decomposition 4.7 Real VectorSpaces 4.8 Problems.......

56 61 63

67 70

76 82 82 87 89 91 93

101 103

109 109 112 114 117 125 129 130 138

IT Infinite-Dimensional Vector Spaces

143

5 Hilbert Spaces 5.1 The Question of Convergence . 5.2 The Space of Square-IntegrableFunctions 5.3 Problems...................

145 145 150 157

6

Generalized Functions 6.1 ContinuousIndex 6.2 Generalized Functions . 6.3 Problems........

159

Classical Orthogonal Polynomials 7.1 General Properties . . 7.2 Classification......... 7.3 Recurrence Relations . . . . 7.4 Examples of Classical Orthogonal Polynomials.

172

7

159 165 169

172 175 176 179

CONTENTS

7.5 7.6 7.7 8

Expansion in Terms of Orthogonal Polynomials .. Generating Functions . . . . . . . . . . . . Problems .

Fourier Analysis

8.1 8.2 8.3

xv 186 190 190 196 196 208 220

Fourier Series . . . . . The Fourier Transform . Problems........

III Complex Analysis

225

9

227 227 228 236 241 248 252 263

Complex Calculus 9.1 Complex Functions " 9.2 Analytic Functions. . . . . . . " 9.3 Conformal Maps .. . . . . . . . . . . . 9A Integration of Complex Functions . . . . . . . . . . . . . " 9.5 Derivativesas Integrals 9.6 Taylor and Laurent Series . 9.7 Problems............................

270 270 273 275 290

10 Calculus of Residues

10.1 10.2 10.3 lOA

Residues . . . . . . . . . . . . . . . . Classificationof Isolated Singularities Evaluation of Definite Integrals . . .. Problems .

11 Complex Analysis: Advanced Topics

ILl 11.2 11.3 1104 11.5 11.6

Meromorphic Functions . . . Multivalued Functions . . . . . . Analytic Continuation . . . . . . The Gamma and Beta Functions. Method of Steepest Descent . . . Problems. . . . . . . . . . . . .

293

. . . . .

. . . . .

293 295 302 309 312 319

. . . .

IV DifferentialEquations

325

12 Separation of Variables in Spherical Coordinates

12.1 12.2 12.3 1204 12.5

PDEs of Mathematical Physics .. . . . . . . . Separation of theAngular Part of the Laplacian. Construction of Eigenvalues of L2. . . . . . . . Eigenvectors of L2: Spherical Harmonics . . Problems.. . . . . . . . . . . . . . ., . . . .

. . . . . . . . . . . . .. . . .

. . .

327 327 331 334 338 346

xvi

CONTENTS

13 Second-Order Linear Differential Equations 13.1 General Properties of ODEs. . . . . . . . . . . . . 13.2 Existence and Uniqneness for First-Order DEs . 13.3 General Properties of SOLDEs . . . 13.4 The Wronskian. . . . . . . . . . . . 13.5 Adjoint Differential Operators. . . . 13.6 Power-Series Solntions of SOLDEs . 13.7 SOLDEs with Constant Coefficients 13.8 The WKB Method . . . . . . . . . . . . 13.9 Numerical Solntions of DEs . . . . . . . 13.10 Problems. . . . . . . . . . . . . . . . . . . . . . . . . . .

348 349 350 352 355 364 367 376 380 383 394

14 Complex Aualysis of SOLDEs 14.1 Analytic Properties of Complex DEs 14.2 Complex SOLDEs . . . . . . . . . . 14.3 Fuchsian Differential Equations . . . 14.4 The Hypergeometric Functiou . . . . 14.5 Confiuent Hypergeometric Functions 14.6 Problems. . . . . . . . . . . . . . . . . .

400 401 404 410 413 419 426

15 Integral Transforms and Differential Equations 15.1 Integral Representation of the Hypergeometric Function . . 15.2 Integral Representation of the Confiuent Hypergeometric Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.3 Integral Representation of Bessel Functions 15.4 Asymptotic Behavior of Bessel Functions. 15.5 Problems.. . . . . . . . . . . . . . . . . . . . . . . . . ..

433 434 437 438 443 445

V Operators on Hilbert Spaces

449

16 An Inlroduction to Operator Theory 16.1 From Abstract to Integral and Differential Operators . 16.2 Bounded Operators in Hilbert Spaces . . 16.3 Spectra of Linear Operators . . 16.4 Compact Sets . 16.5 Compact Operators . 16.6 Spectrum of Compact Operators 16.7 Spectral Theorem for Compact Operators . 16.8 Resolvents 16.9 Problems. . . . . . . . . . . . . . . . . .

451 451 453 457 458 464 467 473 480 485

17 Integral Equations 17.1 Classification.

488 488

CONTENTS

17.2 Fredholm Integra!Equations . 17.3 Problems............

xvii

494 505

18 Sturm-Liouville Systems: Formalism 18.1 Unbounded Operators with Compact Resolvent 18.2 Sturm-Liouville Systems and SOLDEs . . . . 18.3 Other Properties of Sturm-Liouville Systems . 18.4 Problems. . . . . . . . . . . . . . . . . . . .

507 507 513 517 522

19 Sturm-Lionville Systems: Examples 19.1 Expansions in Termsof Eigenfunctions . 19.2 Separation in Cartesian Coordinates. . . 19.3 Separation in Cylindrical Coordinates. . 19.4 Separation in Spherical Coordinates. 19.5 Problems.................

524 524 526 535 540 545

VI Green's Functions

551

20 Green's Functions in One Dimension 20.1 Calculationof Some Green's Functions . 20.2 Formal Considerations . . . . . . . . . . 20.3 Green's Functions for SOLDOs . . . . . 20.4 EigenfunctionExpansion of Green's Fnnctions . 20.5 Problems.. . . . . . . . . . . . . . . . . . . . . . . . . . .

553 554 557 565 577 580

21 Multidimensional Green's Functions: Formalism 21.1 Properties of Partial DifferentialEquations . 21.2 MultidimensionalGFs and Delta Functions . 21.3 Formal Development . . . 21.4 Integral Equations and GFs 21.5 PerturbationTheory 21.6 Problems. . . . . . . . . .

583 584 592 596 600 603 610

22 Multidimensional Green's Functions: Applications 22.1 Elliptic Equations . . 22.2 Parabolic Equations . . . . . . . . . . . . . . . 22.3 HyperbolicEquations . . . . . . . . . . . . . . 22.4 The Fourier TransformTechnique . . . . 22.5 The EigenfunctionExpansion Technique 22.6 Problems. .. .. .. .. .. .. . ...

613 613 621 626 628 636 641

r-

xviii

CONTENTS

VII

Groups and Manifolds

649

23 Group Theory 23.1 Groups. 23.2 Subgroups . . . 23.3 Group Action . 23.4 The Symmetric Group 23.5 Problems. . . . . . . . .

652 656 663 664 669

24 Group Representation Theory 24.1 Definitions and Examples 24.2 Orthogonality Properties. . . 24.3 Analysis of Representations . . 24.4 Group Algebra . . . . . . . .. . . . . . 24.5 Relationship of Characters to Those of a Subgroup . 24.6 Irreducible Basis Functions . . . . . . . 24.7 Tensor Product of Representations . . . 24.8 Representations of the Symmetric Group 24.9 Problems.. .. .. .. . .. .. .. ..

673 680 685 687 692 695 699 707 723

25 Algebra of Tensors 25.1 Multilinear Mappings 25.2 Symmetries of Tensors . 25.3 Exterior Algebra . . . . . 25.4 hmer Product Revisited . 25.5 The Hodge Star Operator 25.6 Problems. . . . . . . . .

729 736 739 749 756 758

26 Analysis of Tensors 26.1 Differentiable Manifolds. . 26.2 Curves and Tangent Vectors . 26.3 Differential of a Map ... 26.4 Tensor Fields on Manifolds 26.5 Exterior Calculus ... 26.6 Symplectic Geometry 26.7 Problems. . . . . . . .

763 770 776 780 791 801 808

s,

VIII

Lie Groups and Their Applications

27 Lie Groups and Lie Algebras 27.1 Lie Groups and Their Algebras . . . . . . . 27.2 An Outline of Lie Algebra Theory ... 27.3 Representation of Compact Lie Groups . . .

651

673

728

763

813 815 815 833 845

-J

CONTENTS

27.4 Representationof the General Linear Group . 27.5 Representationof Lie Algebras 27.6 Problems .

xix

856 859 876

28 DifferentialGeometry 28.1 VectorFields and Curvature . . . . . . 28.2 Riemannian Manifolds. . . . 28.3 CovariantDerivative and Geodesics . 28.4 Isometries and Killing VectorFields 28.5 Geodesic Deviation and Curvature . 28.6 General Theory of Relativity 28.7 Problems. . . . . . . . . . . . . . . . . . . . . . . . . . . .

882 883 887 897 908 913 918 932

29 Lie Gronps and DifferentialEquations

936 936 941 951 956 964 970

29.1 29.2 29.3 29.4 29.5 29.6

Synunetries of Algebraic Equations. . . . . . Synunetry Groups of Differential Equations The Central Theorems. . . . . . . Application to Some Known PDEs Application to ODEs. Problems. . . . . . . . . . . . . .

30 Calcnlusof Variations, Symmetries, and ConservationLaws 30.1 The Calculus of Variations. . . . . . . . . . . 30.2 Symmetry Groups of VariationalProblems . . . . . . . . 30.3 ConservationLaws and Noether's Theorem. . . . . . . . 30.4 Application to Classical Field Theory . 30.5 Problems. . . . . . . . . . . . . . . . . . . . . . . . . . ..

973 973 988 992 997 1000

Bibliography

1003

Index

1007

List of Symbols

E, (11')

z R

R+ C l\I

ill ~A

AxB An

U,

«»

A=B x ..... f(x) V 3 [a] gof

iff ek(a, b) Cn(or JRn) pClt] P'[t] P~[t]

Coo (al b)

lIall

"belongs to", ("does not belong to") Set of integers Set of real nnmbers Set of positive real nnmbers Set of complex nnmbers Set of nonnegative integers Set of rational nnmbers Complement of the set A Set of ordered pairs (a, b) with a E A and b E B {(aI, a2, ... , an)lat E

Al

Union, (Intersection) A is eqnivalent to B x is mapped to f(x) via the map f for all (valnes of) There exists (a valne of) Eqnivalence class to which a belongs Composition of maps f and g if and only if Set of functions on (a, b) with continnons derivatives np to order k Set of complex (or real) n-tnples Set of polynomials in t with complex coefficients Set of polynomials in t with real coefficients Set of polynomials with complex coefficients of degree n or less Set of all complex seqnences (atl~1 snch that L:~I latl2 < 00 Inner prodnct of la) and Ibl Norm (length) of the vector la)

xxii

LIST OF SYMBOLS

£-(V) [8, T] Tt At, or A

l.lE/lV 8(x - xo)

Res[f(zolJ DE, ODE, PDE SOLDE GL(V) GL(n,q SL(n,q 71 ®72

AJ\B AP(V)

Set of endomorphisms (linear operators) on vector space V Commutator of operators 8 and T Adjoint (hermitian conjugate) of operator T Transpose of matrix A Direct sum of vector spaces 1.l and V Dirac delta function nonvanishing only at x = xo Residue of f at point zo Differential equation, Ordinary DE, Partial DE Second order linear (ordinary) differential equation Set of all invertible operators on vector space V Set of all n x n complex matrices of nonzero determinant Set of all n x n complex matrices of unit determinant Tensor product of 71 and 72 Exterior (wedge) product of skew-symmetric tensors A and B Set of all skew-symmetric tensors of type (p, 0) on V

0

_

Mathematical Preliminaries

This introductory chapter gathers together some of the most basic tools and notions that are used throughout the book. It also introduces some common vocabulary and notations used in modem mathematical physics literature. Readers familiar

with such concepts as sets, maps, equivalence relations, and metric spaces may wish to skip this chapter.

0.1. Sets

concept ofset elaborated

Modem mathematics starts with the basic (and undefinable) concept of set. We think of a set as a structureless family, or collection, of objects. We speak, for example, of the set of students in a college, of men in a city, of women working for a corporation, of vectors in space, of points in a plane, or of events in the continuum of space-time. Each member a of a set A is called an element of that sel. This relation is denoted by a E A (read "a is an element of A" or "a belongs to A"), and its negation by a ¢ A. Sometimes a is called a point of the set A to emphasize a geometric connotation. A set is usually designated by enumeration of its elements between braces. For example, {2, 4, 6, 8} represents the set consisting of the first four even natural numbers; {O, ±I, ±2, ±3, ... } is the set of all integers; {I, x, x 2 , x 3, ... } is the set of all nonnegative powers of x; and {I, i, -1, -i} is the set of the four complex fourth roots of unity. In many cases, a set is defined by a (mathematical) statement that holds for all ofits elements. Such a set is generally denoted by {x IP (x)} and read "the set of all x's such that P(x) is true." The foregoing examples of sets can be written alternatively as follows: {n I n is even and I < n < 9}

2

O. MATHEMATICAL PRELIMINARIES

{±n I n is a natural number} {y I y = x" and n is a uatural uumber} {z I Z4 = I and z is a complex uumber} In a frequently used shorthand uotation, the last two sets can be abbreviated as [x" I n 2: and n is an integer} and [z E iC I Z4 = I}. Similarly, the uuit circle can be deuoted by {z [z] = I}, the closed interval [a, b] as {xla ::; x ::; b}, the open interval (a, b) as {x I a < x < b}, and the set of all nonnegative powers of x as {x"}~o' This last notation will be used frequeutly iu this book. A set with a single element is called a singleton. If a E A whenever a E B, we say that B is a subset of A and write B C A or A:J B. If Be A and A c B, then A = B.1f Be A and A i' B, thenB is called a proper subset of A. The set defined by {ala i' a}is called the empty set and is denoted by 0. Clearly, 0 contains no elements and is a subset of any arbitrary set. The collection of all subsets (including 0) of a set A is denoted by 2 A • The reason for this notation is that the number of subsets of a set containing n elements is 2" (Problem a.I).1f A and B are sets, their union, denoted by A U B, is the set containing all elements that belong to A or B or both. The intersection of the sets .!\ and B, denoted by A n B, is the set containing all elements belonging to both A and B. If {B.}.El is a collection of sets, 1 we denote their union by U.B. and their intersection by n. BaIn any application of set theory there is an underlying universal set whose subsets are the objects of study. This universal set is usually clear from the context. For exaunple, in the study of the properties of integers, the set of integers, denoted by Z, is the universal set. The set of reals, JR, is the universal set in real analysis, and the set of complex numbers, iC, is the universal set in complex analysis. With a universal set X in mind, one can write X ~ A instead of ~ A. The complement of a set A is denoted by ~ A and defined as

a

singleton

(proper) subset empty set

union, intersection, complement

universal set

~

A sa {a I a

Ii! A}.

The complement of B in A (or their difference) is A ~B

Cartesian product ordered pairs

== {ala

E

A and a

Ii! B}.

From two given sets A and B, it is possible to form the Cartesian prodnct of A and B, denoted by A x B, which is the set of ordered pairs (a, b), where a E A and b E B. This is expressed in set-theoretic notatiou as A x B

= {(a, b)la E A

and b e B}.

1Here I is an index set--or a counting set-with its typical element denoted by ct. In most cases, I is the set of (nonnegative) integers, but, in principle, it can be any set, for example, the set of real numbers.

0.1 SETS

3

We can generalize this to an arbitrary number of sets. If AI, Az, ... , An are sets, then the Cartesian product of these sets is

Al x Az x ... x An = {(ai, az, ... , an)!ai E Ad, which is a set of ordered n-tuples. If Al An instead of A x A x··· x A, and

= Az = ... = An = A, then we write

An = {(ai, az, ... , an) I a; E Aj. The most familiar example of a Cartesian product occurs when A = R Then JRz is the set of pairs (XI,xz) with XI, xz E JR. This is simply the points in the Euclidean plane. Similarly, JR3 is the set of triplets (XI,xz, X3), or the points in space, and JRn = {(XI, Xz, ... , Xn)!Xi E JRj is the set of real n-tuples.

0.1.1 Equivalence Relations There are many instances in which the elements of a set are naturally grouped together. For example, all vector potentials that differ by the gradient of a scalar function can be grouped together because they all give the same magnetic field. Similarly, all quantum state functions (of unit "length") that differ by a multiplicative complex number of unit length can be grouped together because they all represent the same physical state. The abstraction of these ideas is summarized in the following definition. relation and equivalence relation

0.1.1. Definition. Let A be a set. A relation on A is a comparison test between ordered pairs ofelements of A. If the pair (a, b) E A x A pass this test, we write at> b and read "a is related to b" An equivalence relation an A is a relation that has the fallowing properties: af>a

a s-b

~

bs a

a i-b.b» c ==>-

equivalence class

(reflexivity)

V'aEA,

(symmetry)

a.b e A, a[>c

a.b;c

E

A,

(transivity)

When a t> b, we say that "a is equivalent to b" The set [a] = {b E Alb t> aj ofall elements that are equivalent to a is called the equivalence class ofa. The reader may verify the following property of equivalence relations.

0.1.2. Proposition. If » is an equivalence relation an A and a, b E A, then either [a] n [b] = 0 or [a] = [bl representative ofan equivalence class

Therefore, a' E [a] implies that [a'] = [a]. In other words, any element of an equivalence class can be chosen to be a representative of that class. Because of the symmetry of equivalence relations, sometimes we denote them by c-o.

4

O. MATHEMATICAL PRELIMINARIES

0.1.3. Example. Let A be the set of humanbeings.Let a »b be interpretedas "a is older than b." Then clearly, I> is a relation but not an equivalence relation. On the other hand, if we interpret a E> b as "a and b have the same paternal grandfather," then l> is an equivalence relation, as the reader may check. The equivalence class of a is the set of all grandchildren of a 's paternal grandfather. Let V be the set of vector potentials. Write A l> A' if A - A' = V f for some function

f. The reader may verify that" is an equivalence relation. and that [A] is the set of all vector potentials giving rise to the same magnetic field. Let the underlying set be Z x (Z - {OJ). Say "(a, b) is related to (c, d)" if ad = be. Then this relation is an equivalence relation. Furthermore, [(a, b)] can be identified as the ratio a/b. l1li

partition ofa set

0.1.4. Definition. Let A be a set and {Ra} a collection ofsubsets ofA. We say that {Ra} is a partition of A, or {Ra} partitions A, if the Ra' s are disjoint, i.e., have

no elementin common, and Uo;Ba = A. quotient set

Now consider the collection {[a] I a E A} of all equivalence classes of A. These classes are disjoint, aod evidently their union covers all of A. Therefore, the collection of equivalence classes of A is a partition of A. This collection is denoted by A/1><1 aod is called the quotient set of A under the equivalence relation 1><1.

0.1.5. Example. Let the underlyingset be lll.3. Define an equivalence relation on lll.3 by saying that PI E lR3 and P2 E }R3 are equivalent if they lie on the same line passing through

projective space

the origin. Then ]R3 I l>
Z/" = {[O], [1], ... , [k - I]}, as the reader is urged to verify. For the equivalence relation defined on IE x IE of Example 0.1.3, the set IE x lEI be identified with lQ, the set of rational numbers.

l>
II

0.2 Maps To communicate between sets, one introduces the concept of a map. A map map, domain, codomain, image

f

from a set X to a set Y, denoted by f : X -> Y or X ~ Y, is a correspondence between elements of X aod those of Y in which all the elements of X participate,

2Purthermore, we need to identify any two points on the edge of the hemisphere which lie on the same diameter.

0.2 MAPS

5

Figure 1 The map f maps all of the set X onto a subset of Y. The shaded area in Y is f(K), the range of f.

and each element of X corresponds to only one element of Y (see Figure 1). If y E Y is the element that corresponds to x E X via the map f, we write y = f(x)

function

or

x

f-->

f(x)

or

and call f (x) the image of x onder f. Thus, by the definition of map, x E X can have only one image. The set X is called the domain, and Y the codomain or the target space. Two maps f : X --> Y and g : X --> Y are said to be equal if f(x) = g(x) for all x E X.

0.2.1. Box. A map whose codomain is the set ofreal numbers IR or the set

ofcomplex numbers iC is commonly called a function.

identity map

A special map that applies to all sets A is idA : A --> A, called the identity map of A, and defined by

VaEA.

graph ofa map

The graph r f of a map

rf

preimage

f : A --> B is a subset of Ax B defined by

= {(a, f(a)) Ia E A} C A x B.

This general definition reduces to the ordinary graphs encountered in algebra and calculus where A = B = IR and A x B is the xy-plane. If A is a subset of X, we call f(A) = {f(x)lx E A} the image of A. Similarly, if B C f(X), we call f- 1(B) = {x E Xlf(x) E B) the inverse image, or preimage, of B. In words, f- 1 (B ) consists of all elements in X whose images are in B C Y.1f B consists of a single element b, then (b) = {x E Xlf(x) = b) consists of all elements of X that are mapped to b. Note that it is possible for many points of X to have the same image in Y. The subset f(X) of the codomain of a map f is called the range of f (see Figure 1).

r:'

6

O. MATHEMATICAL PRELIMINARIES

Figure 2

composition oftwo maps

Thecomposition of two maps is another map.

If I : X --> Y and g : Y --> W, then the mapping h : X --> W given by h(x) = g(f(x)) is' called the composition of I and g, and is denoted by h = g 0 I (see Figure 2).3 It is easy to verify that loidx=l=idyol

injection, surjection, and bijection, or 1-1 correspondence

inverse of a map

If l(xI) = I(X2) implies that XI = X2, we call I injective, or one-to-one (denoted I-I). For an injective map only one element of X corresponds to an element of Y. If I(X) = Y, the mapping is said to be surjective, oronto. A map that is both injective and surjective is said to be bijective, or to be a one-toone correspondence. Two sets that are in one-to-one correspondence, have, by definition, the same nnmber of elements. If I : X --> Y is a bijection from X onto Y, then for each y E Y there is one and only one element X in X for which I(x) = y. Thus, there is a mapping I-I : Y --> X given by I-I(y) = x, where X is the nniqne element such that I(x) = y. This mapping is called the inverse of I. The inverse of I is also identified as the map that satisfies I 0 I-I = id y and I-I 0 I = id x- For example, one can easily verify that ln-I = exp and exp"! = ln, because In(e X ) = X and elnx = x. Given a map I : X --> Y, we can define a relation txl on X by saying XI txl X2 if l(xI) = I(X2). The reader may check that this is in fact an equivalence relation. The equivalence classes are subsets of X all of whose elements map to the same point in Y. In fact, [x] = 1-1(f(X». Corresponding to I, there is a map! : X/txl--> Y given by !([x]) = I(x). This map is injective because if !([XI]) = !([X2]), then l(xI) = I(X2), so XI and X2 belong to the same equivalence class; therefore, [XI] = [X2]. It follows that! : X/txl--> I(X) is bijective. If I and g are both bijections with inverses I-I and g -I, respectively,then go I also has an inverse, and verifying that (g 0 f)-I = I-log-I is straightforward.

3Note the importance of the order in whichthe composition is written. The reverse ordermaynot even exist.

0.3 METRIC SPACES

7

0.2.2. Example. As an example of the preirnage of a set, consider the sine and cosine

functions. Then it should be clearthat . -10

SID

=

{nn:JOO n=-oo'

1

cos- 0 =

{i +

mT:

J:-

oo

Similarly, sin- 1[O,

!l consists of all the intervals on the x-axis marked by heavy line segments in Figure 3, i.e., all thepointswhose sine lies between0 and ~. III

in{ectivity and surjectivity depend on the domain and codomain

As examples of maps, we consider fonctions 1 : lR -+ lR stndied in calculus. The Iwo fonctions 1 : lR -+ lR aud g : lR -+ (-I, +I) given, respectively, by 1 (x) = x 3 aud g(x) = taubx are bijective. The latter function, by the way, shows that there are as mauy points in the whole real line as there are in the interval (-I, +1). If we denote the set of positive real numbers by lR+, then the function 1 : lR -+ lR+ given by I(x) = x 2 is surjective but not injective (both x aud ~x map to x 2 ). The function g : lR+ -+ lR given by the same rule, g(x) = x 2 , is injective but not surjective. On the other haud, h : lR+ -+ lR+ again given by h(x) = x 2 is bijective, but u : lR -+ lRgiven by the same rule is neither injective nor surjective.

unit circle

binary operation

LetM n x n denote the set ofn xn real matrices. Define a function det : M n x n ---+ lR by det(A) = det A,where det Ais the determinaut ofAfor A E J\1nxn. This fonction is clearly surjective (why?) but not injective. The set of all matrices whose determinaut is 1 is det- I ( I ). Such matrices occur frequently in physical applications. Another example of interest is 1 : C -+ lRgiven by 1 (z) = [z]. This function is also neither injective nor swjective. Here 1- 1 (1) is the unit circle, the circle of radius I in the complex plaue. The domain of a map cau be a Cartesiau product of a set, as in 1 : X x X -+ Y. Two specific cases are worthy of mention. The first is when Y = R An example of this case is the dot product onvectors, Thus, if X is the set of vectors in space, we cau define I(a, b) = a· b. The second case is when Y = X. Then 1 is called a binary operation on X, whereby au element in X is associated with two' elements in X. For instance, let X = Z, the set of all integers; then the fonction I: Z xZ -+ Zdefined by [tm, n) = mn is the binary operation of multiplication of integers. Similarly, g : lR x lR -+ lR given by g(x, y) = x + y is the binary operation of addition of real numbers.

0.3 Metric Spaces Although sets are at the root of modem mathematics, they are only of formal aud abstract interest by themselves. To make sets useful, it is necessary to introduce some structnres on them. There are two general procedures for the implementation of such structnres. These are the abstractions of the two major brauches of mathematics-algebra aud aualysis.

8

O. MATHEMATICAL PRELIMINARIES

Figure 3 Theunionof all the intervals on the x-axis marked by heavyline segments is . -1[0 ,~1] . sm

We can turn a set into an algebraic structure by introducing a binary operation on it. For example, a vector space consists, among other things, of the binary operation of vector addition. A group is, among other things, a set together with the binary operation of "multiplication". There are many other examples of algebraic systems, and they constitute the rich subject of algebra. When analysis, the other branch of mathematics, is abstracted using the concept of sets, it leads to topology, in which the concept of continuity plays a central role. This is also a rich subject with far-reaching implications and applications. We shall not go into any details of these two areas of mathematics. Although some algebraic systems will be discussed and the ideas of limit and continuity will be used in the sequel, this will be done in an intuitive fashion, by introducing and employing the concepts when they are needed. On the other hand, some general concepts will be introduced when they require minimum prerequisites. One of these is a metric space:

metric space defined

0.3.1. Definition. A metric space is a set X together with a real-valued function d: X x X ~ lR such that

(a) d(x,y) 2: 0 VX,y,andd(x,y)=Oiffx=y. (b) d(x, y) = d(y, x). (c) d(x, y) ::: d(x, z)

(symmetry)

+ d(z, y).

(the triangle inequality)

It is worthwhile to point out that X is a completely arbitrary set and needs no other structure. In this respect Definition 0.3.1 is very broad and encompasses many different situations, as the following examples will show. Before exantining the examples, note that the function d defined above is the abstraction ofthe notion of distance: (a) says that the distance between any two points is always nonnegative and is zero ouly if the two points coincide; (b) says that the distance between two points does not change if the two points are interchanged; (c) states the known fact

0.3 METRIC SPACES

9

that the sum of the lengths of two sides of a triangle is always greater than or equal to the length of the third side. Now consider these examples:

1. Let X = iQI, the set of rational numbers, and define d(x, y) = Ix 2. Let X = R, and again define d(x, y) = Ix -

yl.

yl.

3. Let X consist of the points on the surface of a sphere. We can define two distance functions on X. Let dt (P, Q) be the length of the chord joining P and Q on the sphere. We can also define another metric, dz(P, Q), as the length of the arc of the great circle passing through points P and Q on the surface of the sphere. It is not hard to convince oneselfthat dl and da satisfy all the properties of a metric function. 4. Let eO[a, b] denote the set of continuous real-valued functions on the closed interval [a, b]. We can define d(f, g) = If(x) - g(x)1 dx for f, g E eO(a, b).

J:

5. Let en(a, b) denote the set of bounded continuons real-valned fnnctions on the closed interval [a, b]. We then define

d(f, g) = max lIf(x) - g(x)IJ xe[a,b]

for f, g E eB(a, b). This notation says: Take the absolute valne of the difference in f andg at all x in the interval [a, b] and then pick the maximum of all these values.

The metric function creates a natural setting in whichto test the "closeness" of points in a metric space. One occasionon whichthe ideaof closenessbecomes sequence defined

convergence defined

essential is in the study of a seqnence. A sequence is a mapping s : N --* X from the set of natural numbers N into the metric space X. Such a mapping associates with a positive integer n apoints(n) of the metric space X. It is customary to write Sn (or X n to match the symbol X) instead of s(n) and to enumerate the values of the function by writing {xnJ~I' Knowledge of the behavior of a sequence for large values of n is of fundamental importance. In particular, it is important to know whether a sequence approaches a finite value as n increases.

0.3.2. Box. Suppose thatfor some x andfor any positive real number e, there exists a natural number N such that dtx-; x) < e whenevern > N. Then we say that the sequence {xn}~l converges to x and write limn~oo d(x n l x) = Oar d(xno x) --* 0 or simply X n --* X. It may not be possible to test directly for the convergence of a given sequence because this requires a knowledge of the limit point x. However, it is possible to

10

O. MATHEMATICAL PRELIMINARIES

Figure 4 smaller.

Cauchy sequence

complete metric space

The distance between the elements of a Cauchy sequence gets smaller and

do the next best thing-to see whether the poiots of the sequence get closer and closer as n gets larger and larger. A Cauchy sequence is a sequence for which limm.n-->oo d(x m, xn) = 0, as shown in Figure 4. We can test directly whether or not a sequence is Cauchy. However, the fact that a sequence is Cauchy does not guarantee that it converges. For example, let the metric space be the set of rational numbers IQI with the metric function d(x, y) = Ix - YI, and consider the sequence {xn}~l where X n = Lk~l (- li+ 1 / k. It is clear that X n is a rational number for any n. Also, to show that IX m - Xn I 4 0 is an exercise in calculus. Thus, the sequence is Cauchy. However, it is probably known to the reader that limn --+ oo X n = In2, which is not a rational number. A metric space io which every Cauchy sequence converges is called a complete metric space. Complete metric spaces playa crucial role in modern analysis. The precediog example shows that IQI is not a complete metric space. However, if the limit poiots of all Cauchy sequences are added to IQI, the resulting space becomes complete. This complete space is, of course, the real number system R It tums out that any iocomplete metric space can be "enlarged" to a complete metric space.

0.4 Cardinality cardinalily

The process of counting is a one-to-one comparison of one set with another. If two sets are io one-to-one correspondence, they are said to have the same cardinality. Two sets with the same cardinality essentially have the same "number" of elements. The set Fn = {I, 2, ... , n} is finite and has cardinality n. Any set from which there is a bijection onto Fn is said to be finite with n elements.

Although some steps had been taken before him in the direction of a definitive theory of sets, the creator of the theoryof sets is considered to be Georg Cantor (1845-1918), who was born in Russia of Danish-Jewish parentage but moved to Germany with his parents.

0.4 CARDINALITY

11

His father urged him to study engineering, and Cantor entered the University of Berlin in 1863 with that intention. There he came under the influence of Weierstrass and turned to pure mathematics. He became Privatdozent at Halle in 1869 and professor in 1879. When

he was twenty-nine he published his first revolutionary paper on the theory of infinite sets in the Journal fUrMathematik. Although some of its propositions were deemed faulty by the

older mathematicians, its overall originality and brilliance attracted attention. He continued to publish papers on the theory of sets and on transfinitenumbers until 1897. One of Cantor's main concerns was to differentiate among

infinite sets by "size" and, like Balzano before him, he decided that one-to-one correspondence should be the basic principle. In his correspondence with Dedekind in 1873, Cantor posed the question of whether the set of real numbers can be put into oneto-one correspondence with the integers, and some weeks later he answered in the negative. He gave two proofs. The first is more complicated than the second, which is the one most often used today. In 1874 Cantor occupied himself with the equivalence of the points of a line and the points of R" and sought to prove that a one-to-one correspondence between these two sets was impossible. Three years later he proved that there is such a correspondence. He wrote to Dedekind, "I see it but I do not believe it." He later showed that given any set, it is always possible to create a new set, the set of subsets of the given set, whose cardinal number is larger than that of the given set. If 1'::0 is the given set, then the cardinal number of the set of subsets is denoted by 2~o. Cantor proved that 2~O = C, where c is the cardinal number of the continuum; i.e., the set of real numbers. Cantor's work, which resolved age-old problems and reversed much previous thought, could hardly be expected to receive immediate acceptance. His ideas on transfinite ordinal and cardinal numbers aroused the hostility of the powerful Leopold Kronecker, who attacked Cantor's theory savagely over more than a decade, repeatedly preventing Cantor from obtaining a more prominent appointment in Berlin. Though Kronecker died in 1891, his attacks left mathematicians suspicious of Cantor's work. Poincare referred to set theory as an interesting "pathological case." He also predicted that "Later generations will regard [Cantor's] Mengenlehre as 'a disease from which one has recovered." At one time Cantor suffered a nervous breakdown, but resumed work in 1887. Many prominent mathematicians, however, were impressed by the uses to which the new

theoryhadalreadybeenpatin analysis,measuretheory,andtopology. Hilbert spreadCantor's ideas in Germany, and in 1926 said, "No one shall expel us from the paradise which Cantor created for us." He praised Cantor's transfinite arithmetic as "the most astonishing product of mathematical thought, one of the most beautiful realizations of human activity in the domain of the purely intelligible." Bertrand Russell described Cantor's work as "probably the greatest of which the age can boast." The subsequent utility of Cantor's work in formalizing mathematics-a movement largely led by Hilbert-seems at odds with Cantor's Platonic view that the greater importance of his work was in its implications for metaphysics and theology. That his work could be so seainlessly diverted from the goals intended by its creator is strong testimony to its objectivity and craftsmanship.

12

O. MATHEMATICAL PRELIMINARIES

countably infinite

uncountable sets

Cantor set constructed

Now consider the set of natnral numbers N = {l, 2, 3, ... }. If there exists a bijection between a set A and N, then A is said to be countably infinite. Some examples of countably infinite sets are the set of all integers, the set of even natnral numbers, the set of odd natnral numbers, the set of all prime numbers, and the set of energy levels of the bound states of a hydrogen atom. It may seem surprising that a subset (such as the set of all even numbers) can be put into one-to-one correspondence with the full set (the set of all natural numbers); however, this is a property shared by all infinite sets. In fact, sometimes infinite sets are defined as those sets that are in one-to-one correspondence with at least one of their proper subsets. It is also astonishing to discover that there are as many rational numbers as there are natnral numbers. After all, there are infinitely many rational numbers just in the interval (0, I)-or between any two distinct real numbers. Sets that are neither finite nor countably infinite are said to be uncountable. In some sense they are "more infinite" than any countable set. Examples of uncountable sets are the points in the interval (-1, +1), the real numbers, the points in a plane, and the points in space. It can be shown that these sets have the same cardinality: There are as many points in three-dimensional space-the whole universe-as there are in the interval (-I, +1) or in any other finite interval. Cardinality is a very intricate mathematical notion with many surprising results. Consider the interval [0, 1]. Remove the open interval (~, ~) from its middle. This means that the points ~ and ~ will not be removed. From the remaining portion, [0, ~] U [~, 1], remove the two middle thirds; the remaining portion will then be [0, ~] U [~, ~] U [~, ~] U [~, 1]

(see Figure 5). Do this indefinitely. What is the cardinality of the remaining set, which is called the Cantor set? Intuitively we expect hardly anything to be left. We might persuade ourselves into accepting the fact that the number of points remaining is at most infinite but countable. The surprising fact is that the cardinality is that of the continuum! Thus, after removal of infinitely many middle thirds, the set that remains has as many points as the original set!

0.5 Mathematical Induction

induction principle

Many a time it is desirable to make a mathematical statement that is true for all natural numbers. For example, we may want to establish a formula involving an integer parameter that will hold for all positive integers. One encounters this situation when, after experimenting with the first few positive integers, one recognizes a pattern and discovers a formula, and wants to make sure that the formula holds for all natural numbers. For this purpose, one uses mathematical induction. The essence of mathematical induction is stated as follows:

0.5 MATHEMATICAL INDUCTION

13

0-----------------1-----2--

34-Figure 5

TheCantor set after one, two, three, and four "dissections."

0.5.1. Box. Suppose that there is associated with each natural number (positive integer) n a statement Sn. Then Sn is true for every positive integer provided the following two conditions hold: I. S, is true. 2. If Sm is true for some given positive integer m, then Sm+l is also true.

binomial theorem

We illustrate the use of mathematical induction by proving the binomial theorem:

where we have used the notation

(;) '"

k!(mm~ k)!'

(2)

The mathematical statement Sm is Equation (1). We note that SI is trivially true. Now we assume that Sm is true and show that Sm+1 is also true. This means starting with Equation (1) and showing that

14

O. MATHEMATICAL PRELIMINARIES

Then the induction principle ensures that the statement (equation) holds for all

positiveintegers. Multiply both sides of Eqnation (I) by a

+ b to obtain

(a +b)m+l = t (rn)a m- k+ l b k + t (rn)a m- k b k+ 1. k~O k

Now separate the k

= 0 term

k=O

k

from the first sum and the k

= rn term from the

second sum:

letk=j-linthissum

The second sum in the last line involves j. Since this is a dummy index, we can substitute any symbol we please. The choice k is especially useful because then

we can unitethetwo summations. This gives

If we now use

which the reader can easily verify, we finally obtain

inductive definitions

Mathematical induction is also used in defining quantities involving integers. Such definitions are called inductive definitions. For example, inductive definition is used in defining powers: a l = a and am = am-lao

0.6 Problems 0.1. Show that the number of subsets of a set containing n elements is 2" .

0.6 PROBLEMS

15

0.2. Let A, B, and C be sets in a universal set U. Show that (a) A C BandB C CimpliesA C C. (b) AcBiffAnB=AiffAUB=B. (c) A c Band B C C implies (A U B) c C. (d) AU B = (A ~ B) U (A n B) U (B ~ A). Hint: To show the equality of two sets, show that each set is a subset of the other.

0.3. For each n In =

E

Ixlix -

N, let II < n and Ix + II >

~} .

Find UnIn and nnIn.

0.4. Show that a'

E

[a] implies that [a'] = [a].

0.5. Can you define a binary .operatiou of "multiplication" on the set of vectors in space? What about vectors in the plane?

0.6. Show that (f a g)-1 = g-1 a 1-1 when I and g are both bijections. 0.7. Take any two open intervals (a, b) and (c, d), and show that there are as many points in the first as there are in the second, regardless of the size of the intervals. Hint: Find a (linear) algebraic relation between points of the two intervals. Leibniz rule

0.8. Use mathematical induction to derive the Leibniz rule for differentiating a product: n k d n (n) d I dn-kg dx n (f. g) = k -d k d n-k' k=O x x

'E

0.9. Use mathematical induction to derive the following results: n

'E r k=O

r n+ 1 - 1

k

=

r- I

,

tk=

k=O

n(n+ I).

2

Additional Reading 1. Halmos, P. Naive Set Theory, Springer-Verlag, 1974. A classic text on intuitive (as opposed to axiomatic) set theory coveriog all the topics discussed in this chapter and much more. 2. Kelley, J. General Topology, Springer-Verlag, 1985. The introductory chapter of this classic reference is a detailed introduction to set theory and mappings. 3. Simmons, G.lntroduction to Topologyand Modern Analysis, Krieger, 1983. The first chapter of this book covers not only set theory and mappings, but also the Cantor set and the fact that integers are as abundant as rational

numbers.

Part I

_

Finite-Dimensional Vector Spaces

j

j

j

j

j

j

j

1

_

Vectors and Transformations

Two- and three-dimensional vectors-undoubtedly familiar objects to the readercan easily be generalized to higher dimensions. Representing vectors by their

components, one can conceiveof vectors having N components. This is themost immediate generalization of vectors in the plane and in space, and such vectors are called N-dimensional Cartesian vectors. Cartesian vectors are limited in two respects: Their components are real, and their dimensionality is finite. Some applications in physics require the removal of one or both of these limitations. It is therefore convenient to study vectors stripped of any dimensionality or reality of components. Such properties become consequences of more fundamental definitions. Although we will be concentrating on finite-dimensional vector spaces in this part of the book, many of the concepts and examples introduced here apply to infinite-dimensional spaces as well.

1.1 Vector Spaces Let us begin with the definition of an abstract (complex) vector space.'

1.1.1. Definition. A vector space V over
[z), and so on, called vectors, with the following propertiesr:

1. To every pair ofvectors la) and Ib) in V there corresponds a vector la) + Ib), also in V, called the sum oj la) and Ib), such that (a) la) + Ib) = Ib) + la), 1Keep in mindthat C is the set of complexnumbers and R the set of reels. 2Thebra, ( I, andket, I), notation for vectors, invented by Dirac, is veryuseful whendealing with complexvectorspaces. However, it is somewhat clumsyforcertain topicssuchasnann andmetrics andwill therefore be abandoned in thosediscussions.

20

1. VECTORS ANO TRANSFORMATIONS

(b) la} + (Ib) + Ie)) = (Ia)

+ Ib)) + [c),

(c) There exists a unique vector 10} E V, called the zero vector, such that la} + 10) = Ja} for every vector la), (d) To every vector la} E V there corresponds a unique vector - Ja} (also in V) such that la) + (-Ia}) = 10). scalars are numbers

2. To every complex number: a-{llso called a scalar-s-and every vector la) there corresponds a vector a la) in V such that (a) a(f3la}) = (af3) la), (b) J ]a) = la).

3. Multiplication involving vectors and scalars is distributive:

(a) a(la) (b) (a complex VS. real vector space concept offield summarized

+ Ib}) =

+ 13) la} =

ala) ala)

+ alb).

+ f3la).

The vector space defined above is also called a complex vector space. It is possible to replace iC with IR-the set of real numbers-s-in which case the resulting space will be called a real vector space. Real and complex numbers are prototypes of a mathematical structure called field. A field is a set of objects with two binary operations called addition and multiplication. Each operation distributes with respect to the other, and each operation has an identity. The identity for addition is denoted by and is called additive identity. The identity for multiplication is denoted by I and is called multiplicative identity. Furthermore, every element has an additive inverse, and every element except the additive identity has a multiplicative inverse.

°

1.1.2. Example.

SOME VECTOR SPACES

1. 1R is a vector space overthe fieldof real numbers. 2. C is a vector space overthefieldof realnumbers. 3. C is a vector space over the complex numbers. 4. LetV = R and letthefieldof scalars heC. Thisis not avector space, becauseproperty 2 of Definition 1.1.1 is not satisfied: A complexnumber timesa realnumber is not a realnumber and therefore doesnotbelongto V. 5. The set of "arrows" in theplane(or in space) forms a vector space over1R under the parallelogram law of additiooof planar (or spatial)vectors.

3Complex numbers, particularly when they aretreated as variables, areusually denoted by z, and we shall adhere to this convention in Part ITI. However, in the discussionof vectorspaces, we havefoundit moreconvenient to use lower case Greek letters to denotecomplexnumbers as scalars.

1.1 VECTOR SPACES

21

6. Let :Pe[t] be the set of all polynomials with complex coefficients in a variable t. Then pC[t] is a vector space under the ordinary addition of polynomials and the multiplication of a polynomial by a complex number. In this case the zero vector is

the zero polynomial. 7. For a given positive integer n, let P~[t] be the set of all polynomials with complex coefficients of degree less than or equal to n. Again it is easy to verify that P~[t] is a vector space under the usual addition of polynomials and their multiplication by complex scalars. In particular. the sum of two polynomials of degree less than or equal to n is also a polynomial of degree less than or equal to n, and multiplying a polynomial with complex coefficients by a complex number gives another polynomial of the same type. Here the zero polynomial is the zero vector. 8. The set ~[t] of polynomials of degree less than or equal to n with real coefficients is a vector space over the reals, but it is not a vector space over the complex numbers. 9. Let en consist of all complex a-tuples such as la) = ("1, "2, ... , "n) and Ib) = (!!I, fJ2, ... , fJn)· Let" be a complex number. Then we define la)

+ Ib) =

("t + fJI, "'2 + fJ2,"" "n ala} = (ceq, aa2, ...• aan),

+ fJn),

10) = (0,0, ... , 0), -Ia) = (-"10 -"2,···, -"n)·

n-dimensional complex coordinate space

en

It is easy to verify that is a vector space over the complex numbers. It is called the n-dimensional complex coordinate space. 10. The set of all real n-tuples lRn is a vector space over the real numbers under the It is called the n-dimensional real coordinate space, operations similar to that of or Cartesian n-space. It is not a vector space over the complex numbers.

en.

n-dimensional reai coordinate space, or

Cartesian n-space

11. The set of all complex matrices withm rows andn columns JY[mxn is a vector space under the usual addition of matrices and multiplication by complex numbers. The zero vector is the m x n matrix with all entries equal to zero. 12. Let Coo be the set of all complex sequences la} = {a; }~l such that L~l la;12 < 00. One can show that with addition and scalar multiplication defined componentwise, Coo is a vector space over the complex numbers.

13. The set of all complex-valued functions of a single real variable that are continuous in the real interval (a, b) is a vector space over the complex numbers.

en (a, b) on (a, b) of all real-valued functions of a single real variable that possess continuous derivatives of all orders up to n forms a vector space over the reals.

14. The set

15. The set eOO(a, b) of all real-valued functions on (a, b) of a single real variable that possess derivatives of all orders forms a vector space over the reals.

II

It is ciear from the example above that the existence nf a vector space depends as much on the nature of the vectors as nn the nature of the scalars.

linear independence defined linear combination of

vectors

1.1.3. Definition. The vectors lal), la2), ... , Ian}, are said to be linearly independent iffor "i EC, the relation L:f=t a; lai) = 0 implies a; = 0 for all i. The sum L:f=l a; lai) is called a linear combination of{lai )}f=l'

22

1. VECTORS AND TRANSFORMATIONS

subspace

1.1.4. Definition. A subspace W of a vector space V is a nonempty subset of V with the property that iJla} , Ib) E W, then ala} + fJ Ib) also belongs to W for all a, fJ E C.

A subspace is a vector space in its own right. The reader may verify that the The intersection of two subspaces is intersection of two subspaces is also a subspace. also a subspace. 1.1.5. Theorem. If S is any nonempty set ofvectors in a vector space V, then the span ofa subset ofa set W s of all linear combinations of vectors in S is a subspace of V. We say that vector space Ws is the span of S, or that S spans W S, or that Ws is spanned by S. Ws is sometimes denoted by Span{S}. The proof of Theorem 1.1.5 is left as Problem 1.8. basis defined

1.1.6. Definition. A basis ofa vector space V is a set B of linearly independent vectors that spans all ofV. A vector space that has afinite basis is calledjinitedimensional; otherwise, it is injinite-dimensional.

We state the following without proof (see [Axle 96, page 31]): 1.1.7. Theorem. All bases of a given finite-dimensional vector space have the same number oflinearly independent vectors. This number is called the dimension ofthe vector space. A vector space ofdimension N is sometimes denoted by VN.

components ofa vector in a basis

If 10) is a vector in an N-dimensional vector space V and B = {Ia,} }~1 a basis in that space, then by the definitionof a basis, there exists a nnique (see Problem 1.4) set of scalars {aI, a2,··., an} such that 10) = L~l a, 10,). The set {ail~l is called the components of la} with respect to the basis B. 1.1.8. Example. Thefollowingaresubspaces of someof thevectorspacesconsidered in Example 1.1.2.

• The "space" of real numbers is a subspace of C overthe reals,

• 1R is not a subspace of C overthecomplexnumbers, becauseasexplained in Example 1.1.2,1R cannot be a vector spaceoverthecomplex numbers. • The set of all vectors along a given line going through the origin is a subspace of arrows in theplane(or space)over JR. • P~[t] is a subspace ofpC[t]. • Cn- 1 is a subspace of en when en- 1 is identified withall complex a-tuples with zerolastentry. In general, em is a subspace of en for m < n when em is identified

withall a-tuples whoselast n - m elements arezero. • :MY XS is a subspace of:Mm xn for r :::: m or s ~ n. Here, we identify an r x s matrix withan m x n matrix whoselast m - r rows andn - s columns areall zero. • :P~ [t] is a subspace of p~ [t] for m < n. • P~[t] is a subspace ofP~[t] for m < n. Note that both P~[t] and P~n[t] are vector

spaces Over thereals only.

1.2 INNER PRODUCT

23

• IRm is a subspace of R" form < n. Therefore,lR2• the plane, is a subspace oflR3~ the Euclidean space. Also, R 1

space IR3•

== IRis a subspace of both the plane]R2 and the Euclidean II

1.1.9. Example. The following are bases for the vector spaces given in Example 1.1.2. • The number 1 is a basis for JR, which is therefore one-dimensional. • The numbers 1 and i = .J=I are basis vectors for the vector space Cover R Thus, this space is two-dimensional.

• The number 1 is a basis for Cover C, and the space is one-dimensional. Note that although the vectors are the same as in the preceding item, changing the nature of the scalars changes the dimensionality of the space.

• The set {ex,ey, cz} of the unit vectors in the directions of the three axes forms a basis in space. The space is three-dimensional. • A basis of pert] can be fanned by the monomials 1, t, (2, .... It is clear that this space is infinite-dimensional. standard basis of

en

el, e2, ... ,en,

e

• A basis of en is given by where j is an n-tuple that has a 1 at the jth position and zeros everywhere else. This basis is called the standard basis of en. Clearly, the space has n dimensions. • A basis of JY[mxn is given by 811, e12, ... , eij, ... , emn• where 8U is the m x n matrix with zeros everywhere except at the intersection ofthe ithrow and jth column, where it has a one. • A set consisting of the monomials 1, t, t 2 , ••• , t n fOnTIS a basis of:P~[t]. Thus, this space is (n

+ 1)-dimensional.

en

• The standard basis of is a basis of R" as well. It is also called the standard basis oflllff • Thus, IRn is n-dimensional.

• If we assume that a < 0 < b, then the set of monomials 1, x, x 2 , ... forms a basis for e""(a, b), because, by Taylor's theorem, any fuuction belonging to e""(a, b) can be expanded in an infinite power series about x = O.Thus, this space is infinitedimensional. ..

Given a space V with a basis B = {lai)}i~l' the span of any m vectors (m < n) of B is an m-dimensional subspace of V.

1.2 Inner Product A vector space, as given by Definition 1.1.1, is too general and structureless to be of much physical interest. One useful structure introduced on a vector space is a scalar product. Recall that the scalar (dot) product of vectors in the plane or in space is a rule that associates with two vectors a and b, a real number. This association, denoted symbolically by g : V x V --> E, with g(a, b) = a . b, is symmetric: g(a, b) = g(b, a), is linear in the first (and by symmetry, in the second) factor:"

g(aa + tJb, c) = ag(a, c)

+ tJg(b, c)

or

4A function that is linear in both of its arguments is called a bilinear function.

(aa + tJb) . c = aa· c + tJb· c,

24

1. VECTORS AND TRANSFORMATIONS

gives the "length" of a vector: lal 2 = g(a, a) = a . a 2: 0, and ensures that the only vector with zero length'' is the zero vector: g(a, a) = 0 if and only if a = O. We want to generalize these properties to abstract vector spaces for which the scalars are complex numbers. A verbatim generalization of the foregoing properties, however, leads to a contradiction. Using the linearity in both arguments and anonzero la), we obtain g(i la), i la»

= i 2g(la} , la» = -g(la) , la}).

(1.1)

Either the right-hand side (RHS) or left-hand side (LHS) of this equation must be negative! But this is inconsistent with the positivity of the "length" of a vector, which requires g(la) , la» to be positive for all nonzero vectors, including i la}. The source of the problem is the linearity in both arguments. If we can change this property in such a way that one of the i's in Equation (1.1) comes out complexconjugated, the problem may go away. This requires linearity in one argument and complex-conjugate linearity in the other. Which argument is to be complexconjugate linear is a matter of convention. We choose the first argument to be so.6 We thus have g(a la)

Dirac "bra," ( I, and "kef' I ), notation Is used for Inner products.

inner product defined

+ fJ Ib), Ie}) = a*g(la) , Ie}) + fJ*g(lb) , Ic}),

where a* denotes the complex conjugate. Consistency then requires us to change the symmetry property as well. In fact, we must demand that g(la} , Ib» = (g(lb) , la»)*, from which the reality of g(la) , la})-anecessary condition for its positivity-follows innnediately. The question of the existence of an inner product on a vector space is a deep problem in higher analysis. Generally, if an inner product exists, there may be many ways to introduce one on a vector space. However, as we shall see in Section 1.2.4, afinite-dimensional vector space always has an inner product and this inner product is unique." So, for all practical purposes we can speak of the inner product on a finite-diroensional vector space, and as with the two- and three-dimensional cases, we can omit the letter g and use a notation that involves only the vectors. There are several such notations in use, but the one that will be employed in this book is the Dirac bra(c)ket notation, whereby g(la} , Ib}) is denoted by (al b). Using this notation, we have 1.2.1. Definition. The inner product a/two vectors, la) and Ib), in a vector space (a I b) E
V is a complex number, 1. (alb) = (bla)* 2. (al

(fJ Ib) + Y Ie}) = fJ (al

b)

+ y (al c)

SInourpresent discussion, we areavoiding situations in which a nonzero vectorcan havezero "length." Suchoccasionsarise in relativity, andwe shalldiscussthem in Part VII. 6In some books,particularly in the mathematical literature. the secondargument is chosen to be linear. 7TIris uniqueness holdsup to a certain equivalence of inner products thatwe shallnot get intohere.

1.2 INNER PRODUCT

3. (ala) ,,::0, and positive definite, or Riemannian inner product

sesquilinear

(ala) =0

ifandonlyif

25

ja) = 10).

The last relation is called the positive definite property ofthe inner product. 8 A positive definite inner product is also called a Riemannian inner product, otherwise it is called pseudo-Riemannian.

Note that linearity in the first argument is absent, because, as explained earlier, it would be inconsistent with the first property, which expresses the "symmetry" of the inner product. The extra operation of complex conjugation renders the true linearity in the second argument impossible. Because of this complex conjugation, the inner product on a complex vector space is not truly bilinear; it is commonly called sesqnilinear. A shorthand notation will be useful when dealing with the inner product of a linear combination of vectors. 1.2.2. Box. We write the illS ofthe second equation in the definition above as (alPb + vc). This has the advantage of treating a linear combination as a single vector. The second property then states that if the complex scalars happen to be in a ket, they "split out" unaffected: (alpb

+ vc) = P (al b) + V (al c) .

(1.2)

On the other hand, if the complex scalars happen to be in the first factor (the bra), then they should be conjugated when they are "split out": (Pb

+ vcla) = P* (bl a) + V* (cl a) .

(1.3)

A vector space V on which an inner product is defined is called an inner product space. As mentioned above, all finite-dimensional vector spaces can be turned into inner product spaces.

1.2.3.Example. In this example we introduce some of the most common inner products. The reader is urged to verify that in all cases, we indeed have an inner product. E cn, with la) define an inner product on

• Let la) , Ib)

natural inner product forC"

=

(ej , a2, ... , an) and Ib)

en as

=

(fJj, f32, ... , f3n), and

n

(alb) =aif31 +a2f32+ .. ·+a~f3n = I>if3j· i=l

That this product satisfies all the required properties of an inner product is easily checked. For example, if Ib) = la), we obtain (al a) = lal1 2 + la21 2 +... + la,,1 2, which is clearly nonnegative. 8The positive definiteness must be relaxed in the space-time of relativity theory, in which nonzero vectors can have zero "length."

26

1. VECTORS ANO TRANSFORMATIONS

• Similarly, for la} • Ib) E JRn the same definition (without the complex conjugation) satisfies all the properties of an inner product.

• For Ie} ,Ib} E
weight function of an inner product defined in terms of integrals

E

:Pe[t], the space of all polynomials in t with complex coefficients.

Define

L b

(xl y) es

w(t)x*(t)y(t)dt,

(t.4)

where a and b are real numbers--or infinity-for which the integral exists, and wet) is a real-valued, continuous function that is always strictlypositive in the interval (a, b). Then Equation (1.4) defines an inner product. Depending on the so-called weight function w(t), there can be many different inner products defined on the infinite-dimensional space pC[t].

naturai inner product for complex functions

• Let

I, g

E


L b

UI g)

'"

w(x)!*(x)g(x) dx.

It is easily shown that (II g) satisfies alt the requirements of the inner product if, as in the previous case, the weight function w(x) is always positive in the interval (a, b). This is called the standard inner product on
1.2.1 Orthogonality The vectors of analytic geometry and calculus are often expressed in terms of unit vectors along the axes, i.e., vectors that are of unit length and perpendicular to one another. Such vectors are also important in abstract inner product spaces.

orthogonalilydefined orthonormal basis

1.2.4, Definition. Vectors la) ,Ib) E V are orthogonal if (al b) =0. A normal vector,or normalized vector, Ie} is onefor whicb (e] e) = 1. A basis B = [lei)}~1 in an N-dimensional vector space V is an orthonormal basis if

I

(e,,I eJ), -8"- 'J = { 0 Kronecker delta

ifi =j, if i 'I i.

where 8ij, defined by the last equality, is called the Kronecker delta. 1.2.5. Example. Here are examples of orthonormal bases: • The standard basis of R" (or C"}

leI} = (1,0, ... ,0), le2) = (0, I, .... ,0), ... , len} = (0,0, ... ,1) is orthonormal under the usual inner product of those spaces.

(1.5)

1.2 INNER PRODUCT

Ia,)-l =1£'1)

laz)

!az)

27

\

'-

~ Ie2) \

\

~I~) I~) ~ ~"'a,)11 \------"'1~) Ie,) '- '\

(a)

(b)

\

(d)

(e)

Figure 1.1 The essence of the Gram-Schmidt process is neatly illustrated by the process in two dimensions. This figure, depicts the stages of the construction of two orthonormal vectors.

• Lellek)

= e ikx /v'2n' be functions in iC(O, 2".) withw(x) = 1. Then

(ekl ek)

= -1

2".

fozn e-tkXe,kx . . dx = 1, 0

and for I i' k, (ell ek) =

~ f21r e-ilxikx dx = ~ f2tr ei(k-l)x dx 2".k

Thns, (etlek) =

=

2".k

o. III

~lk'

1.2.2 The Gram-Schmidt Process

The Gram-Schmidt process explained

It is always possible to convert any basis in V into an orthonormal basis. A process by wbich this may be accomplished is called Gram-Schmidt orthonormalization. Consider a basis B = {lal} , laz} , ... , IaN)}. We intend to take linear combinations of lai) in such a way that the resulting vectors are orthonormal. First, we let leI) = lal) /v'(atlal) and note that (ell el) = 1. If we subtract from laz) its projection along leI), we obtain a vector that is orthogonal to leI) (see Figure 1.1). Ca1ling the resulting vector lez)' we have lez) = laz} - (ell az) let), which can be written more symmetrically as 102) = laz) - leI) (ell az). Clearly, this vector is orthogonal to Iet}.In order to normalize lez}, we divide it by

(ezl ez). Then

lez} == lez) / (ezl ez) will be a normal vector orthogonal to leI). Subtracting from la3) its projections along the first and second unit vectors obtained so far will give the vector

z

Ie;) = la3) -leI} (ella3) -lez} (ezla3) = la3) -

L lei} (eil a3) , i=l

28

1. VECTORS ANO TRANSFORMATIONS

la,)

(c)

(b)

(a)

Figure 1.2 Once the orthonormal vectors in the plane of two vectors are obtained, the third orthonormal vector is easily constructed.

which is orthogonal to both lei) and le2) (see Fignre 1.2):

I

=1

~O

,..-'-,

,..-'-,

(ell e3) = (ell a3) - (eli el) (ell a3) - (ell e2) (e21 a3) = O. Similarly, (e21 e~)

= O.

Erhard Schmidt (1876-1959) obtained bis doctorate nnder the supervision of David Hilbert. His main interest was in integral equations and Hilbert spaces. He is the "Schmidt" of the Gram-8chmldtorthogonalization process, which takes a basis of a space and constructs an orthonormal one from it. (Laplace had presented a special case of this process long before Gram or Schntidt.) In 1908 Schmidt worked on infinitely many equations in

infinitelymany unknowns, introducingvariousgeometricnotations and terms that are still in use for describing spaces of functions. Schmidt's ideas were to lead to the geometry of Hilbert spaces. This was motivated by the study of integral equations (see Chapter 17) and an attempt at their abstraction. Earlier, Hilbert regarded a function as given by its Fourier coefficients. These satisfy the condition that L~l a~ is finite. He introduced sequences of real numbers {xn} such that L~l x~ is finite. Riesz and Fischer showed that there is a one-to-one correspondence between square-integrable functions and square-summable sequences of their Fourier coefficients. In 1907 Schmidt and Frechet showed that a consistent theory could be obtained if the square-summable sequences were regarded as the coordinates of points in an infinitedimensional space that is a generalization of n-dimensional Euclidean space. Thusftmctions can be regarded as points ofa space, now called a Hilbert space.

1.2 INNER PRODUCT

29

In general, if we have calculated m orthonormal vectors lell, ... , !em), with m < N, then we can find the next one using the following relatious: m

le~+l) = lam+!) - Lie,) (eilam+, ) , i=l

!e~+l)

(1.6)

(e~+,1 e~'+l) Even though we have been discussing finite-dimensional vector spaces, the process ofEquation (1.6) can continue for infinite-dimensions as well. The reader is asked to pay attention to the fact that, at eacb stage of the Gram-Schmidt process, one is taking linear combinations of the original vectors.

1.2.3 The Schwarz Inequality Let us now consider an important inequality that is valid in both finite and infinite dimensions and whose restriction to two and three dimensions is equivalent to the fact that the cosine of the angle between two vectors is always less than one. . Schwarz inequality

1.2.6. Theorem. For any pair of vectors la) , Ib) in an inner product space V, the Schwarz inequality holds: (al a) (bl b) 2: 1 (al b) 12 . Equality holds when la) is proportional to Ib). Proof Let [c) = Ib) - ((al b) / (al a)) la), and note that (al c) = O. Write Ib) = ((al b) / (al a») la) + Ic) and take the inner product of Ib) with itself: (bib) =

l

b)12 (ala) + (clc).

(al (ala)

Since the last term is nevernegative, we have 2

(bl b) > -

I

(al b) 1 (al a)

=}

(al a) (bl b) 2: 1 (al b) 12 •

Equality holds iff (c] c) = 0 or Ic) = O. From the definition of [c), we conclude that la) and Ib) must be proportional. 0 Notice the power of abstraction: Webave derived the Schwarz inequality solely from the basic assumptions of inner product spaces independent of the specific nature of the inner product. Therefore, we do not have to prove the Schwarz inequality every time we encounter a new inner product space.

Karl Herman Amandus Schwarz (1843-1921) the son of an architect, was born in what is nowSobiecin, Poland. Aftergymnasium. Schwarz studied chemistry in Berlin fora time

30

1. VECTORS ANO TRANSFORMATIONS

before switching to mathematics, receiving his doctorate in 1864. He was greatly influenced by the reigning mathematicians in Germany at the time, especially Kummerand Weierstrass. The lecture notes that Schwan took while attending Weierstrass's lectures on the integral calculus still exist. Schwarz received an initial appointment at Halle and later appointments in Zurich and Gotringen before being named as Weierstrass'g successor at Berlin in 1892. These later years, filled with students and lectures, were not Schwarz's most productive,

but his early papers assure his place in mathematicshistory. Schwarz's favorite tool was geometry, which he soon turned to the study of analysis.He conclusively proved some of Riemann's results that had been previously (and justifiably) challenged. The primary result in question was the assertion that every simply connected region in the plane could be conformally mapped onto a circular area. From this effort came several well-known results now associated with Schwarz's name, including the principle of reflection and Schwarz's lemma. He also worked on surfaces of minimal area, the

brancbof geometrybelovedby all who dabblewith soap bubbles. Schwarz's most important work, for the occasion of Weierstrass's seventieth birthday, again dealt with minimal area, specifically whether a minimal surface yields a minimal area. Along the way, Schwarz demonstrated second variation in a multiple integral, constructed a function using successive approximation, and demonstrated the existence of a "least" eigenvalue for certain differential. equations. This work also contained the most famous inequality in mathematics, which bears his name. Schwarz's success obviously stemmed from a matching of his aptitude and training to the mathematical problems of the day. One of his traits, however, could be viewed as either positive or negative-his habit of treating all problems, whether trivial or monumental, with the same level of attention to detail. This might also at least partly explain the decline in productivity in Schwarz's later years. Schwarz had interests outside mathematics, although his marriage was a mathematical one, since he married Kummer's daughter. Outside mathematics he was the captain of the local voluntary fire brigade, and he assisted the stationmaster at the local railway station by closing the doors of the trains!

1.2.4 Length of a Vector In dealing with objects such as directed line segments in the plane or in space, the intuitive idea of the length of a vector is used to define the dot product. However, sometimes it is more convenient to introduce the inner product first and then define the length, as we shall do now. norm of avector defined

1.2.7. Definition. The norm, or length, ofa vector la) in an inner product space is denoted by lIall and defined as lIall ==.J"(£il
31

1.2 INNER PROOUCT

1. The norm of the zero vector is zero: 11011 = 0. 2. lIall 2:: 0, and lIall = 0 if and only if la) = 10). 3. lIaall = lailiall tor any? complex o. triangle inequality narmed linear space

natural distance in a normed linear space

parallelogram law

4. lIa + bll ::0 lIall

+ IIbli. This property is called the triangle ineqnality.

A vector space on which a nonn is defined is called a normed linear space. One can introduce the idea of the "distance" between two vectors in a nonned linear space. The distance between la) and Ib)----
then one can define

and show that it is indeed an inner product. In fact, we have (see [Frie 82, pp. 203-204] for a proof) the foUowing theorem.

1.2.8. Theorem. A normed linear space is an inner product space the norm satisfies the parallelogram law.

if and only if

Now consider any N-dimensional vector space V. Choose a basis

Ilain!:!

in

V, and for any vector la) whose components are la; 1!:1 in this basis, define N

lIall

2

== ~)ail2. i=l

The reader may check that this defines a norm, and that the norm satisfies the parallelogrann law. From Theorem 1.2.8 we have the foUowing:

1.2.9. Theorem. Everyfinite-dimensional vector space can be turned into an inner product space.

32

1. VECTORS ANO TRANSFORMATIONS


1.2.10. Example. Let the space be en. The natural inner product of en gives rise to a an) is

norm, which,forthevector la} = (cq , aZ

I ••• ,

functions n

lIall=~=

~)a;l2. i=l

This norm yields thefollowing distance between la) and Ib) = Uh.lh. ... . fJn): n

d(a,b) =

Iia-bil =)(a-bla-b)=

~)ai -fJiI 2 • i=l

Onecandefine nthernorms, such as lIaIIJ sa LI~l lai I,which hasall thereqnired properties of a nonn, andleadsto the distance n

dl (a. b) = lIa - bill = ~)aj -

fJ;I·

i=1

Another norm defined on en is given by lI ali p

=

(~Ia;lp ) lip

where p is a positive integer. It is proved inhigher mathematical analysis that II . lip hasall theproperties of anonn.(Thenontrivial part of theproofis to verifythe triangle inequality.) Theassociated distance is dp(a, b)

= lIa -

blip

("

= \;~ laj -

fJ;lP

)tl P

Theother two noons introduced above arespecialcases, for p = 2 andp = 1.

III

1.3 Linear Transformations We have made progress in enriching vector spaces with structures such as norms and inner products. However,this enrichment, although important, will be of little value if it is imprisoned in a single vector space. We would like to give vector space properties freedom of movement, so they can go from one space to another. The vehicle that carries these properties is a linear transformation which is the subject of this section. However, first it is instructive to review the concept of a mapping (discussed in Chapter 0) by considering some examples relevant. to the present discussion.

1.3.1. Example. Thefollowing area few familiar examples of mappings. 9Thefirstproperty follows fromthisby letting a

=

O.

1.3 LINEAR TRANSFORMATIONS

33

1. Let f : IR ->-1R be given by f(x) = x 2 •

2. Let g : 1R2 ->- IR be given by g(x, y) = x 2 + y2 - 4. 3. Let F : 1R2 ->- iC be given by F(x, y) = U(x, y) + iV(x, y), where U : 1R2 ->- IR and V : 1R2 ->- lR. 4. Let T : IR ->- 1R2 be given by T(t) = (t + 3, 2t - 5). 5. Motion ofapointparticle in space can be considered as a mapping M : [a, b] --+ JR3 , where [a, b] is an interval of the real line. For each t E [a, b], we define M(t) = (x(t), y(t), z(t», where x(t), y(t), and z(t) are real-valned functions of t. If we identify t with time, which is assumed to have a value in the interval [a, b], then M(t) describes the path of the particle as a function of time, and a and b arethe beginning and theendof themotion, respectively. II Let us consider an arbitrary mapping F : V ---> W from a vector space V to another vector space W. It is assumed that the two vector spaces are over the same scalars, say Co Consider la) and Ib) in V and Ix) and Iy) in W such that F(la) = Ix) and F(lb) = Iy). In general, F does not preserve the vector space structnre. That is, the image of a linear combination of vectors is not the same as

the linearcombination of the images: F(a la)

+ {J Ib) i' aF(lx) + {JF(ly).

This is the case for all the mappings of Example 1.3.1 except the fourth item. There are many applications in which the preservation of the vector space structure (preservation of the linear combination) is desired. linear transformation, linear operator, endomorphism

1.3.2. Definition. A linear transformation from the complex vector space V to the complex vector space W is a mapping T : V ---> W such that T(a la) + {J Ib)) = aTria))

+ {JT(lb))

Via) , Ib) E V and a, {J E C.

A linear transformation T : V ---> V is called an endomorphism of V or a linear operator on V. The action ofa linear transformation on a vector is written without the parentheses: T(la)) sa T la). The same definition applies to real vector spaces. Note that the definition demands that both vector spaces have the same set of scalars: The same scalars multiply vectors in V on the LHS and those in W on the RHS. An immediate consequence of the definition is the following:

1.3.3. Box. Twa linear transformations T : V ---> Wand U : V ---> Ware equal if and only ifT la,) = U lail for all la, ) in same basis afV. Thus, a linear transformation is uniquely determined by its action an same basis of its domain space.

34

1. VECTORS AND TRANSFORMATIONS

linear functional

qv, W) isavector space

dual vector space v·

The equality in this box is simply the set-theoretic equality of maps discussed in Chapter O. Au importaut example of a linear trausformation occurs when the second vector space, W, happens to be the set of scalars, C or JR, in which case the linear trausformation is called a linear functional. The set of linear trausformations from V to W is denoted by r:. (V, W), aud this set happens to be a vector space. The zero transformation, 0, is defined to take every vector in V to the zero vector of W. The sum of two linear trausformations T aud U is the lineartrausformation T+U, whose action on a vector la) E V is defined to be (T+ U) la) es T la) + U la). Similarly, define aT by (aD la) es arT tal) = aT la). The set of endomorphisms of V is denoted by r:. (V) rather thau t: (V, V). The set oflinear functionals r:. (V, iC)--or t: (V, JR) if V is a real vector space--is denoted by V* aud is called the dual space of V. 1.3.4. Example. Thefollowing aresomeexamples of linear operators in various vector spaces.Theproofsof linearity are simplein all cases andareleft as exercises forthereader. I. Let {Jal} , la2) , ... , lam}} be an arbitraryfinite set of vectorsin V, and {I), 12, ... , 1m } an arbitrary set of linear functionals on V. Let m

A'"

L

lak)lk E qV)

k~l

be defined by A [x) = L:~llak) Ik(lx» operator on V.

=

L:~1 Ik(lx» lak}. Then A is a linear

2. Let" be a permutation (shnffiing) of the integers It, 2, ... , n}. If [x} = (~1, ~2, ... , fJn) is a vector in en, we can write An Ix} = (~n(l)' ~n(2), ... , ~n(n).

ThenAn is a linear operator.

derivative operator integration operator

integration is a linear functional on the space of continuous functions

3. For any Ix) E :P'[t], withx(t) = L:Z~o cxktk, write Iy} = D [x}, where Iy) is defined as yet) = Lk=l kaktk-l. Then Dis a linear operator, thederivativeoperator.

with x(t) = L:Z~o cxktk, write Iy) = S Ix}, where Iy) E :P'[t] is defined as yet) = L:Z~O[CXk/(k + l)]tH I. Then S is a linear operator, the integrationoperator.

4. For every [z}

E :P'[t],

5. Define the operator Inl : eO(a, b) -+ IRby

inl(f) =

l

b

f(t) dt.

Thenintis a linear functional on thevectorspace eO(a, b)~

6. Let en(a, b) be the set ofreat-valued functions definedin the interval [a, b] whose first n derivatives exist and are continoous. For any If) E en(a, b) define lu} = G If}, with u(t) = g(t)f(t) and get) a fixed function in en (a, b). Then G is linear. In particular. the operation of multiplying by t, whose operator is denoted by T, is ~~

.

1.3 LINEAR TRANSFORMATIONS

35

An immediate consequence of Definition 1.3.2 is that the image of the zero vector in V is the zero vector in W. This is not true for a general mapping, but it is necessarily true for a linear mapping. As the zero vector of V is mapped onto the zero vector of W, other vectors of V may also be dragged along. In fact, we have the following theorem.

kernel ofa linear transformation

1.3.5. Theorem. The set ofvectors in V that are mapped onto the zero vector of W under the linear transformation T : V --+ W form a subspace of V called the Tremel, or null space, ofT and denoted by kerT.

D

Proof The proof is left as an exercise". nullity

rank ofa linear transformation

The dimension of ker T is also called the nullity of V. The proof of the following is also left as an exercise. 1.3.6. Theorem. The range T(V) of a linear transformation T subspace ofW. The dimension ofT(V) is called the rank ofT. 1.3.7. Theorem. A linear transformation is I-I (injective)

V--+Wisa

iff its kernel is zero.

Proof The "only if" part is trivial. For the "if" part, suppose T lat) = T la2); then linearity of T implies that T(lat) - la2)) = O. Since kerT = 0, we must have lat) = la2). D Suppose we start with a basis of ker T and add enough linearly independent vectors to it to get a basis for V. Without loss ofgenerality, let us assume that the first n vectors in this basis form a basis ofker T. So let B = (Ial) , la2) , ... , IaN)} be a basis for V and B' = (Iat) , la2) , ... , Ian)} be a basis for kerT. Here N = dim V andn = dim kerT. II is straightforward to show that (Tlan+t), ... , T IaN)} is a basis for T(V). We therefore have the following result. dimension theorem

1.3.8. Theorem. Let T : V --+ W be a linear transformation. Then!O dim V = dim ker T + dim T(V) This theorem is called the dimension theorem. One ofits consequences is that an injective endomorphism is automatically surjective, and vice versa: 1.3.9. Proposition. An endomorphism of a finite-dimensional vector space is bijective zf it is either injective or surjective. The dimension theorem is obviously valid only for finite-dimensional vector spaces. In particnlar, neither swjectivity nor injectivity implies bijectivity for infinite-dimensional vector spaces.

lORecall that the dimension of a vector space depends on the scalars used in that space. Although we are dealing with two different vector spaces here, since they are both over the same set of scalars (complex or real), no confusion in the concept of dimension arises.

36

1. VECTORS AND TRANSFORMATIONS

1.3.10. Example. Let us try to find the kernel ofT : ]R4 -+

]R3

givenby

2xl +x2 +x3 -X4 =0, XI +x2 +2X3 +2X4 = 0, XI - X3 - 3X4 = 0.

The"solution" to theseequations is Xl = kerT, a vector in must be of the form

r

x3

+ 3X4 and X2

= -3X3 - 5X4.

Thus, to be in

where X3 andX4 are arbitrary realnumbers. It follows that kerTconsists of vectors that can be written as linearcombinations of the two linearly independent vectors (1, -3, 1, 0) and (3, -5,0, I). Therefore, dim kerT = 2. Theorem 1.3.8then says that dimT(V) = 2; that is, therange of T is two-dimensional. Thisbecomesclearwhen one notesthat

andtherefore T(x!. X2. X3, X4), anarbitrary vectorin therange ofT, is a linear combination of only two linearlyindependentvectors, (I, 0, I) and (0, I, -I). III

In many cases, two vector spaces may "look" different, while in reality they are very much the same. For example, the set of complex numbers C is a twodimensional vector space over the reals, as is 1R2. Although we call the vectors of these two spaces by different names, they have very similar properties. This notion of "similarity" is made precise in the following definition. isomorphism and automorphism

1.3.11. Definition. A vector space V is said to be isomorphic to another vector space W if there exists a bijective linear mapping T ; '17---> W. Then T is called an isomorphism. II A bijective linear map of V onto itselfis called an automorphism ofV. The set ofautomarphisms of V is denoted by GL(V). For all practical purposes, two isomorphic vector spaces are different manifestations of the"same" vectorspace. Intheexample discussed above, thecorrespondence T; C ---> 1R2 , with T(x + iy) = (x, y), establishes an isomorphism between the two vectorspaces. It should be emphasized that only as vector spaces are C and 1R2 isomorphic. If we go beyond the vector space structures, the two sets are qnite different. For example, C has a natural multiplication for its elements, but 1R2 does not. The following theorem gives a working criterion for isomorphism.

llThe word "isomorphism," as we shall see, is used in conjunction with many algebraic structures. To distinguish them, qualifiers need to be used. In the present context, we speak of linear isomorphism. We shall use qualifiers when necessary. However, the contextusuallymakesthe meaning of isomorphism clear.

1.3 LINEAR TRANSFORMATIONS

37

1.3.12. Theorem. A linear surjective map T : V --+ W is an isomorphism only if its nullity is zero.

if and

Proof The "only if" part is obvious. To prove the "if" part, assume that the nnllity is zero. Then by Theorem 1.3.7, T is I-I. Since it is already surjective, T must be bijective. D 1.3.13. Theorem. An isomorphism T : V --+ W carries linearly independent sets ofvectors onto linearly independent sets ofvectors.

Proof Assume that {Iai) }I~I is a set of linearly independent vectors in V. To show that{T lai))I=1 is linearly independent in W, assume that there exist eq , az, ... , am such that 2:1~1 a; T lai) = 10). Then the linearity of T and Theorem 1.3.12 give T(2:7:"1 a; lai)) = 10), or 2:;'::1 a; laj) = 10), and the linear independence of for all i. Therefore, {T laj ))I~I must be linearly the laj) implies that a, = independent. D

°

The following theorem shows that finite-dimensional vector spaces are severely limited in number: 1.3.14. Theorem. Twofinite-dimensional vector spaces are isomorphic if and only

if they have the same dimension. Proof Let Bv = {lai)}~1 be a basis for V and Bw = {lbi))~1 a basis for W. Define the linear transformation Tlai) = Ibi), i = 1,2, ... , N. The rest of the proof involves showing that T is an isomorphism. We leave this as an exercise for the reader. D only two N-dimensional vector spaces

A consequence of Theorem 1.3.14 is that all N-dimensional vector spaces over JR are isomorphic to JRN and all complex N-dimensional vector spaces are isomorphic to eN. So, for all practical purposes, we have only two N-dimensional vector spaces, IRN and eN.

1.3.1 More on Linear Functionals An example of isomorphism is that between a vector space and its dual space, which we discuss now. Consider an N-dimensional vector space with a basis B = {Ial), laz), ... , ION)). For any given set of N scalars, {ai, az, ... , aN}, define the linear functional fa by fa laj) = at. When fa acts on any arbitrary vector Ib) = 2:~dJ; laj) in V, the result is (1.9) This expression suggests that Ib) can be represented as a column vector with entries th, fJz, ... , fJN and fa as a row vector with entries ai, az,···, aN. Then fa Ib) is

38

1. VECTORS ANO TRANSFORMATIONS

Every set of N scalars defines a linear functional.

merely the matrix product'< of the row vector (on the left) and the column vector (on the right). fa is uniquely detennioed by the set {aI, a2, ... ,aN j. In other words, corresponding to every set of N scalars there exists a unique linear functional. This leads us to a particular set of functionals, fl, f2, , fN corresponding, respectively, to the sets of scalars {I, 0, 0, ... , OJ, {O, 1,0, , OJ, .,., (O, 0, 0, ... , I], This means that fl 101) = 1

and

fllaj} =0

for

j

f21a2) = 1

and

f2laj} =0

for

j

i' 1, i' 2,

for

j

i' N,

and or that

(1.10) where 8ij is the Kronecker delta. The functionals of Equation (1.10) form a basis of the dual space V*. To show this, consider an arbitrary 9 E V*, which is uniquely determined by its action on the vectors in a basis B = lIal) , la2} , ... , ION} j. Let 9 lai} = Yi E C. Then we claim that 9 = L:;:"I Yifi. In fact, consider an arbitrary vector 10) in V with components (aI, a2, ... , aN) with respect to B. Then, on the one hand, 9 10) = g(L:;:"1 a; lai) = L:;:"I aig lai} = L:;:"! aiYi. On the other hand,

(~Yifi) la} = (~Yifi) (~aj laj)) N N = LYi Lajfi laj) ;=1

j=l

N

N

N

= LYi LaAj = LYiai. ;=1

j=l

i=l

Since the actions of 9 and L:;:"I Yifi yield equal results for arbitrary 10), we conclude that 9 = L:;:"I Yifi, i.e., {fi j;:"1 span V'. Thus, we have the following result. 1.3.15. Theorem. If V is an N-dimensional vector space with a basis B = lIal} , 102) , ... , ION} j, then there is a corresponding unique basis B* = {f;};:"! in V* with the property that fi 10 j} = 8ij.

dual basis

By this theorem the dual space of an N -dinnensional vector space is also Ndinnensional, and thus isomorphic to it. The basis B* is called the dual basis of B. A corollary to Theorem 1.3.15 is that to every vector in V there corresponds a

12Matrices will be taken up in Chapter 3. Here,we assumeonly a nodding familiarity with elementary matrix operations.

1.3 liNEAR TRANSFORMATIONS

39

unique linear functional in V*. This can be seen by noting that every vector la) is uniquely determined by its components (ai, a2, ... ,aN) in a basis B. The unique linear functional fa corresponding to [a), also called the dual of la), is simply

Lf::l ajfj, with Ii E B*. annihilator ofa vector and a subspace

1.3.16. Definition. An annihilator of la) E V is a linear junctional f E V* such that f la) = O. Let W be a subspace ofV. The set of linear functionals in V* that annihilate all vectors in W is denoted by WO. The reader may check that WO is a subspace of V*. Moreover, if we extend a basis {Iai))f~t of W to a basis B = (Iaj) 1[:,1 of V, then we can show that the functionals {f j 17=k+l' chosen from the basis B* = (f j 17=t dnalto B, span WO. I! then follows that dim V = dimW+dimWO.

(1.11)

We shall have occasions to use annihilators later on when we discuss symplectic geometry. We have "dualed" a vector, a basis, and a complete vector space. The only object remaining is a linear transformation. dual, or pull back, of a linear transformation

1.3.17. Definition. Let T : V --> 11 be a linear transformation. Define l ' : 11* --> V* by13 [T*(g)J la) = g(T la»)

Via) E V, 9 E 11*,

T* is called the dual or pull back, ofT. One can readily verify that l ' E £'(11*, VOl, i.e., that T* is a linear operator on 11*. Some of the mapping properties of T* are tied to those of T. To see this we first consider the kernel of 1'. Clearly, 9 is in the kernel of l ' if and only if 9 annihilates all vectors of the form T [a), i.e., all vectors in T(V). I! follows that 9 is in T(V)o. In particular, ifT is surjective, T(V) = 11,and 9 annihilates all vectors in 11, i.e., it is the zero linear functional. We conclude that ker T* = 0, and therefore, T* is injective. Similarly, one can show that if T is injective, then T* is surjective. We summarize the discussion above:

1.3.18. Proposition. Let T be a linear transformation and T* its pull back. Then kerT* = T(V)o. 1fT is surjective (injective), then T* is injective (surjective). In particular, T" is an isomorphism ifT is.

duals and inner products

I! is useful to make a connection between the inner product and linear functionals. To do this, consider a basis (Iat) , la2) , ... , IaN)) and letaj = (al aj). As noted earlier, the set of scalars (ail[:,t defines a unique linear functional fa such that t, lai) = ai. Since (a Iaj) is also equal to cq, ilis natural to identify fa with

13Donot confusethis "*,, with complexconjugation.

40

1. VECTORS AND TRANSFORMATIONS

tihe symbol (ai, and write T: fa >--> (al where T is tiheidentification map. It is also convenient to introduce tihe notation'f (Ia))t es (ai,

dagger ofa linear combination of

vectors

(1.12)

where tihe symbol t means "dual, or dagger of." Now we ask: How does tihis dagger operation act on a linear combination of vectors? Let ]c) = ala) + fJ Ih) and take tihe inner product of [c) witih an arbitrary vector Ix} using linearity iu tihe second factor: (xl e) = a (xl a) + fJ (xl h). Now complex conjugate botih sides and use tihe (sesqui)symmetry of tihe inner product: (LHS)* = (xle)* = (cl x}, (RHS)*

= a* (xl a)* + fJ* (xl h)* = a* (al x) + fJ* (hi x) = (a* (al + fJ* (hi) [x) .

Since tihis is true for all [x), we must have (Ie))t es (c] = a* (al+fJ* (hi. Therefore, in a duality "operation" tihecomplex scalars must be conjugated. So, we have (a la)

+ fJ Ih})t =

a* (al

+ fJ* (hi.

(1.13)

Thus, unlike tihe association la) .... la which is linear, tihe association la .... (al is uotlinear, but sesquilinear, i.e., tihe ideutification map T mentioned above is also sesquilinear: T(ala

+ fJlb) = a* (al + fJ* (hi = a*T(la) + fJ*T(lb);

It is convenient to represent lu) E

Compare (1.14) with the comments after (1.9). The complex conjugation in (1.14) isthe resuit ofthe sesquilinearily ofthe association la) <+ (s],

en as a column vector

Then tihe definition of the complex inner product suggests tihattihe dual of la) must be represented as a row vector witihcomplex conjugate entries: (al = (aj

ai

...

a~),

and tihe inner product can be written as tihe (matrix) product

(al h) = (aj

ai

14The significance of thisnotation will becomeclearin Section2.3.

(1.14)

1.4 ALGEBRAS

41

=

Iluj )}i'~l and BV = (IVj}}j=l' respectively. Consider an mn-dimensional vector space W whose basis Bw is in one-to-one correspondence with the pairs (lUi), IVj», and let !UiVj) be the vector corresponding to (Iuj), IVj»' For lu) Ell with components 10I;}i'=1 in Bu and [v) E V with components {11jlj=l in Bv , definethe vector lu, v} E W whosecomponents in Bw

1.3.19. Example. Let IL and V be vector spaces with bases BU

are {OIi~j}' One can easily sbow that if lu), lu'), and lu") are vectors in IL and lu") = 01 lu) + fJ lu'), then [a". v) =

01 lu, v)

+ fJ lu', v)

The space W thus defined is calledthe tensor product oflL and V and denotedby IL <8> V. One can also define tensor productof three or more vector spaces. Of specialinterest are tensor products of a vector space and its dual. The tensor product space Vr,s of type (r, s) is definedas follows:

V,., = 'V<8> v V <8>..." V<8>V* <8>V* <8>, ... <8> V*. I

r times

s times

We sball comeback to this space in Cbapter25.

III

1.4 Algebras In many physical applications, a vector space has a natural "product," the prime example being the vector space of matrices. It is therefore useful to consider vector spaces for which such a product exists. algebra defined

1.4.1. Definition. An algebra A over iC (or IR) is a vector space over iC (or IR), together with a binary operation /L : V x V -+ V, called multiplication, that satisfies 15

a(pb + ye) = pab + yac dimension ofthe algebra; associativity; commutativity; identity; and right and leftinverses

Va, b, e E A, V p, Y E iC (or IR),

with a similar relation for multiplication on the right. The dimension ofthe vector space is called the dimension ofthe algebra. The algebra is called associative if the product satisfies a(bc) = (ab)e and commutative if it satisfies ab = ba. An algebra with identity is an algebra that has an element 1 satisfying al = la = a. An element b ofan algebra with identity is said to be a left inverse of a ifba = 1. Right inverse is defined similarly. 1.4.2. Example. Definethe following producton ]Rz: (xj , XZ)(Yl, yz) = (XIYl - XZYZ, X1YZ

+ XZYl)·

The reader is urged to verify that this product turns JR2 into a commutativealgebra. 15We shall abandon the Dirac bra-and-ket notation in this section due to its clumsiness; instead we use boldface roman letters to denote vectors. It is customary to write ab for p,(a, b).

42

1. VECTORS ANO TRANSFORMATIONS

Similarly, the vector(cross)product on lR.3 turns it into a nonassociative, noncornmu-

tative algebra. The paradigm of all algebras is thematrix algebra whose binary operation is ordinary multiplication of n x n matrices. This algebra is associative butnot commutative. All the examples above are finite-dimensional algebras. An example of an infinitedimensional algebra is eOO(a, b), the vector space of infinitely differentiable real-valued

functions onarealinterval (a, b). Themultiplication is defioed pointwise: H f and g E eOO(a, b), then fg(x) '" f(x)g(x)

vx

E

(a, b).

This algebra is commutative andassociative.

E

eOO(a, b)

.

The last item in the example above has a feature that turns out to be of great significancein all algebras, the product rule for differentiation.

1.4.3.Definition. A vector space endomorphism D : A --+ A is calleda derivation derivation of an algebra defined

on A

if it has the additional property

D(ab) = [D(a)]b + a[D(b)]. 1.4.4. Example. Let A be thesetof n x n matrices. Definethebinary operation, denoted bye, as AoB"'AB-BA,

where the RHS is ordinary matrix multiplication. The reader may check that A together withthis operation becomes an algebra. Now let A be a fixed matrix, anddefinethe linear transformation

Thenwe note that DA(B 0 C) = A. (B. C) = A(B. C) - (B 0 C)A =A~-~-~-C~=~-_-_+~

Ontheother hand, (DAB). C + B. (DAC) = (AoB) .C+B. (AoC) = (AB- BA) .C+B. (AC - CAl = (AB - BA)C- C(AB - BA) + B(AC - CAl - (AC- CA)B = ABC+ CBA- BCA- ACB.

So, DA is a derivation on A.

..

The lineartransformations connectiogvector spacescan be modifiedslightlyto accommodatethebinary operationof multiplicationof the correspondingalgebras:

1.4 ALGEBRAS

algebra homomorphism and isomorphism

43

1.4.5. Definition. Let A and 13 be algebras. A linear transformation T : A ..... 13 is called an algebra homomorphism if T(ab) = T(a)T(b). A bijective algebra homomorphism is called an algebra isomorphism.

1.4.6. Example. LetA be ]R3, and :8 thesetof 3 x 3 matrices of theform

Then themap T : A

-+ :8 defined by

can be shown to be a linear isomorphism. Let the cross product be the binary operation on A, tnming it intoan algebra. For B, define thebinary operation of Example 1.4.4. The reader may checkthat, with theseoperations, Tis extended to an algebra isomorphism. IlIlI

structure constants ofanalgebra

Given an algebra A and a basis B = (e;}~1 for the underlying vector space, one can write N eiej =

Lc~ekl

(1.15)

k~1

The complex numbers efj , the components of the vector eiej in the basis B, are called the structure constants of A. These constants determine the product of any two vectors once they are expressed in terms of the basis vectors of B. Conversely,

1.4.7. Box. Given any N-dimensional vector space V, one can turn it into an algebra by choosing a basis and a set ofN 3 numbers {c~} and defining the product ofbasis vectors by Equation (1.15).

1.4.8. Example. Consider the vector space of n x n matrices with its standard basis where e., hasa 1 at the ijthposition and zero everywhere else. Thismeans that

{ei]'}~l,j_ "-I'

(eij)zk = 8il8jko (eijekl)mn

and n

n

r=1

r=1

= L(eij)mr(ekl)rn = L~im~jr8kr8In =

lJimlJjk8In

= 8jk(eU)mn,

or eijekl = 8jk eil·

The structure constants are cjj~kl these constants.

= 8im8jk8ln. Note thatone needs a double index to label II

44

1. VECTORS AND TRANSFORMATIONS

1.4.9. Example. In the standard basis{eil of R4, choose the structore constants as follows:

2=2 = -02 -e3 = -e42 = ej , et2 elei = eie! = ei for i = 2, 3, 4, eiej

=

L Eijkek

for i, j

= 2, 3, 4,

k

[
algebra of quaternions

is defined in Equation (3.19)]. The reader may verify that these relations turn R4

into an associative, butnoncommutative, algebra, called the algebra of quaternions and

denoted by 1HI. In thiscontext, ej is usually denoted by I, ande2, e3, ande4 by i, j, and k, respectively, and one writes q = x

+ iy + iz + kw for an element oflBI. It thenbecomes

evident that1HI is a generalization of C. In analogy withC, x is called the realpart of q, and (y, z, w) the pure part of q. Similarly, the conjugateof q is q* = x - iy - iz - kw. III Algebras have a surprisingly rich structure and are used extensively in many branches of mathematics and physics. We shall see their usefulness in our discussion of group theory in Part vn. To ciose this section, and to complete this introductory discussion of algebras, we cite a few more useful notions. left, right, end two-sided ideals

1.4.10. Definition. Let A be an algebra. A subspace 'B ofA is called asubalgebra ofA if'B contains the products ofall its members. If'B has the extra property that it contains ab for all a E A and b E 'B, then 'B is called a left ideal of A. A right ideal and a two-sided ideal are defined similarly.

It is ciear from the definition that an ideal is automatically a subalgebra, and that 1.4.11. Box. No proper ideal of an algebra with identity can contain the identity element.

minimal ideal

In fact, no proper left (right) ideal can contain an element that has a left (right) inverse. An ideal can itself contain a proper (sub)ideal. If an ideal does not contain any proper subideal, it is called a minimal ideal. 1.4.12. Example. Thevector space eO(a, b) of all contiuuous real-valued functions ou an interval (a, b) is turned into a commutative algebra by pointwise multiplication: If f, g E eO(a, b), thenthe product fg is defined by fg(x) '" f(x)g(x) for all x E (a, b). The set of functions that vanish at a given fixed point C E (a, b) constitutes an ideal in eO(a, b). Since the algebra is commutative, the ideal is two-sided.

ideals generated by an element ofan algebra

II

One can easily construct left ideals for an algebra A: Take any element x E A and consider the set Ax", {ax I a E A}.

1.5 PROBLEMS

45

The reader may check that .Ax is a left ideal. A right ideal can be constructed similarly. To construct a two-sided ideal, consider the set

.AKA == {axb I a, h E .A}. These are all called ideals generated by x.

1.5 Problems 1.1. Let R+ denote the set of positive real nnmbers. Define the "sum" of two elements of R+ to be their usual product, and define scalar multiplication hy elements of R as being given by r . p = pr where r E R and p E R+. With these operations, show that R+ is a vector space over R 1.2. Show that the intersection of two subspaces is also a subspace.

1.3. For each of the following subsets of R3 determine whether it is a subspace ofR3 : (a) {(X, y, z) ER3 jx + y - 2z = 0}; (b) {(x,y,z)ER 3Ix+y- 2z=3); (c) {(x, y, z) E R3 1 xyz = OJ. 1.4. Prove that the components of a vector in a given basis are unique.

en (or Rn).

1.5. Show that the following vectors form a basis for

I I

I

I

o

I

... ,

laz) = I I

I

o

o o

1.6. Let W be a subspace of R5 defined by W = {(Xl, ... ,X5) E R 5 1xI = 3xz +X3, Xz = X5, and xa = 2x3}. Find a basis for W. 1.7. Show that the inner product of any vector with 10) is zero. 1.8. Prove Theorem 1.1.5. CO, Ci, and cz such that the polynomials aD, bo + bIt, and + Cit + czt Z are mutually orthonormal in the interval [0, I]. The inner product

1.9. Find aD, bo, bl, CO

is as defined for polynomials in Example 1.2.3 with wet) = 1.

46

1. VECTORS AND TRANSFORMATIONS

1.10. Given the linearly independent vectors x(t) = t", for n = 0, 1,2, ... in pc[I], use the Gram-Schmidt process to find the orthonormal polynomials eo(t), ei (t), and e2(t) (a) when the inner product is defined as (xl y) = J~l x*(t)y(t) dt. (b) when the inner product is defined with a nontrivial weight function:

i:

(xl y) =

e-t'x*(t)y(t) dt.

Hint: Use the following result:

..;n

00

ioo

2

e-ttndt=

j

ifn = 0, if n is odd,

0

..;n

1 ·3·5 .. · (n - 1) 2n/ 2

if n is even.

1.11. (a) Use the Gram-Schmidt process to find an orthonormal set of vectors out of (I, -I, 1), (-1,0, I), and (2, -1,2). (b) Are these three vectors linearly independent? If not, find a zero linear combination of them by using part (a). 1.12. (a) Use the Gram-Schmidt process to find an orthonormal set of vectors out of (I, -I, 2), (-2, I, -1), and (-1, -I, 4). (b) Are these three vectors linearly independent? If not, find a zero linear combination of them by using part (a). 1.13. Show that

i: i: i: i:

1.14. Show that dx

dy (x 5 - x 3 + 2x 2 - 2)(y5

s

dx

-l + 2y2 -

dy (x 4 - 2x 2 + l)(i

4

2)e-(x +

y4

)

4

+ 4y3 + 4)e-(x +y' ).

Hint: Define an appropriate inner product and use the Schwarz inequality.

1.15. Show that for any set of n complex numbers ai, a2, ... , an, we have la1 + a2 + ... + an 12 :s n (la112

+ la21 2 + ... + Ian 12) .

Hint: Apply the Schwarz inequality to (I, 1, ... ,1) and (ai, a2, ... , an).

1.5 PROBLEMS

1.16. Using the Schwarz inequality show that if (a;}f,;;, and then Lf,;;, al fJ; is convergent.

ltl;}f,;;,

47

are iu iCoo,

1.17. Show that T : IR2 --> IR3 given byT(x, y) = (x 2 + y2,x + y, 2x - y) is not a linear mapping. 1.18. Verify that all the transformations of Example 1.3.4 are linear. 1.19. Let 17: be the permutation that takes (1,2,3) to (3, 1, 2). Find A" Ie;}, i = 1,2,3, where {Ie; ple 1.3.4.

}}T=, is the standard basis ofIR3 (or i(3) , and A" is as defined in Exam-

1.20. Show that if TEI:,(iC, C), then there exists a E iC such that T la} = ala} for all la) E C. 1.21. Show that if (la;}}f~l spans V and TEI:,(V, W) is surjective, then (T la,}}f=l spans W. 1.22. Give an example of a function

f(a la}) = af(la}) but

f

f : IR2 --> IR such that

Va E IR and la} E IR2

is not linear. Hint: Consider a homogeneous function of degree 1.

1.23. Show that the following transformations are linear: (a) V is ic over the reals and C [z) = [z"). Is C linear if instead of real numbers, complex numbers are used as scalars?

(b) V is pelt] and T Ix(t) = Ix(t + 1) - Ix(t)}. 1.24. Verify that the kernel of a transformation T : V --> W is a subspace of V, and that T(V) is a subspace of W. 1.25. Let V and W be finite dimensional vector spaces. Show thatifTEI:,(V, W) is surjective, then dim W ::: dim V. 1.26. Suppose that V is finite dimensional and TEl:, (V, W) is not zero. Prove that there exists a subspace U of V such that ker T n U = (OJ and T(V) = T(U). 1.27. Show that WO is a subspace of V* and dim V = dimW+ dimWo. 1.28. Show that T and T* have the same rank. In particular, show that if T is injective, then T* is surjective, Hint: Use the dimension theorem for T and T* and Equation (1.11).

48

1. VECTORS ANO TRANSFORMATIONS

1.29. Show that (a) the product on IR2 defiued by

turns IR2 into an associative and commutative algebra, and (b) the cross product on IR.3 turns it intoa nonassociative, noncommutative algebra. 1.30. Fix a vector a E IR3 and define the linear transformation D. : IR3 -4 IR3 by D.(h) = a x b. Show that D. is a derivation ofIR 3 with the cross product as multiplication. 1.31. Show that the linear transformation of Example 1.4.6 is an isomorphism of the two algebras A and 11. 1.32. Write down all the structure constants for the algebra of quaternions. Show that this algebra is associative. 1.33. Show that a quaternion is real iff it commntes with every quaternion and that it is pure iff its square is a nonpositive real number. 1.34. Let p and q be two quaternions. Show that (a) (pq)* = q*p*. (b) q E IRiffq* = q, andq E IR3 iffq* = -q. (c) qq* = q*q is a nonnegative real number. 1.35. Show that no proper left (right) ideal of an algebra with identity can contain an element that has a left (right) inverse. 1.36. Let A be an algebra, and x E A. Show that Ax is a left ideal, xA is a right ideal, and AxA is a two-sided ideal.

Additional Reading I. Alder, S. Linear Algebra Done Right, Springer-Verlag, 1996. A small text packed with information. Lots of marginal notes and historical remarks. 2. Greub, W. Linear Algebra, 4th ed., Springer-Verlag, 1975. Chapter V has a good discussion of algebras and their properties. 3. Halmos, P. Finite Dimensional Vector Spaces, 2nd ed., Van Nostrand, 1958. Another great book by the master expositor.

2

_

Operator Algebra

Recall that a vector space in which one can multiply two vectors to obtain a third vector is called an algebra. In this cbapter, we want to iuvestigate the algebra of linear transformations. We have already established that the set of linear transformations L (V, W) from V to W is a vector space. Let us attempt to define a multiplication as well. The best candidate is the composition of linear transformations. If T : V --+ U and S : U --+ W are linear operators, then the composition SoT: V --+ W is also a linear operator, as can easily be verified. This product, however, is not defined on a single vector space, bnt is such that it takes an element in L(V, U) and another element in a second vector space L(U, W) to give an element in yet another vector space L(V, W). An algebra requires a single vector space. We can accomplish this by letting V = U = W. Then the three spaces oflinear transformations collapse to the single space L (V, V), the set of endomorphisms of V, which we abbreviate as ,c,(V) and to which T, S, ST sa SoT, and TS sa To S belong. The space ,c,(V) is the algebra of the linear operators on V.

2.1 Algebra of 'c(V) Operator algebra encompasses various operations on, and relations among, operators. One of these relations is the equality of operators, which is intuitively obvions; nevertheless, we make it explicit in (see also Box 1.3.3) the following defiuition. operator equality

2.1.1. Definition. Two linearoperators T, U E ,c,(V) areequal ifT 10) = U la)Ior aUla) E V. Because of the linearity of T and U, we have

50

2. OPERATOR ALGEBRA

2.1.2. Box. Two endomorphisms T, U E L(V) are equal ifT lai} = U lai} for aUla,} E B, where B is a basis ofV. Therefore, an endomorphism is uniquely determined by its action on the vectors ofa basis.

The equality of operators can also be established by other, more convenient, methods when an inner product is defined on the vector space. The following two theorems contain the essence of these alternatives. 2.1.3. Theorem. An endomorphism T of an inner product space is 0 if and only if! {bl T la} es (bl Ta) = Ofor aUla) and Ib).

Proof Clearly, ifT = 0 then {bl T la} = O. Conversely, if (bl T la) = 0 for all ]a) and Ib), then, choosing Ib) = T la) = ITa), we obtain (TaITa)=O Via)

*

Tla)=O

Via)

*

T=O

o

by positive definiteness of the inner product.

2.1.4. Theorem. A linear operator T on an inner product space is 0 if and only if (al T la) = Ofor aUla}.

polarization identity

Proof Obviously, ifT = 0, then (al T la) = O. Conversely, choose a vector a la) + fJ Ib), sandwich T between this vector and its dual, and rearrange terms to obtain what is known as the polarization identity: a* fJ (al T Ib)

+ afJ* (bl T la) =

(aa

+ fJbl T laa + fJb) 2

-lal 2 (al T la) - IfJI (bl T Ib). According to the assumption of the theorem, the RHS is zero. Thus, if we let a = fJ = 1 we obtain (al T Ib) + (hi T la) = O. Similarly, with a =1 and fJ = i we get i (al T Ib) - i (bl T la) = O. These two equations give (al T Ib) = 0 for all la) , Ib). By Theorem 2.1.3, T = O. 0 To show that two operators U and T are equal, one can either have them act on an arbitrary vector and show that they give the same result, or one verifies that U- T is the zero operator by means ofone ofthe theorems above. Equivalently, one shows that (al T Ib) = (al U Ib) or (al T la) = (al U la) for all le}, Ib}. In addition to the zero element, which is present in all algebras, L(V) has an identity element, 1, which satisfies the relation 1 la) = la) for all ]«) E V. With 1 in our possession, we can ask whether it is possible to find an operator T- 1 with the property that T- I T = n- I = 1. Generally speaking, ouly bijective mappings have inverses. Therefore, only automorphisms ofa vector space are invertible. 1It is convenient here to use the notation ITa) for T la). This would then allow us to write the dual of the vector as (Tal, emphasizing that it is indeed the bra associated with T la).

2.1 ALGEBRA OF L(V)

51

2.1.5. Example. Let the linear operator T : lit 3 -+ lit3 be defined by T(xl. x2, X3) = (Xl + X2, X2 + x3, xl

+ x3)·

We want to see whetherT is invertible and, if so, find its inverse. T has an inverse if and only if it is bijective. By the comments after Theorem 1.3.8 this is thecase if andonly if T is either surjective or injective. The latter is equivalent to kef T = IO}. But kef T is the set of all vectors satisfying T(XI, x2, X3) = (0,0, 0), or Xl +X2 =0,

X2+X3

=0,

The reader may check thatthe unique solutionto these equations is xi = x2 = x3 = Thus, the only vector belonging to kef Tis the zerovector. Therefore, Thas an inverse. To find r l applyT-IT = 1 to (XI,X2,X3): (xj , X2, X3)

o.

= T-IT(xl. X2, X3) = T-I(XI + X2, X2 + x3, Xl + X3)·

This equationdemonstrates how T- 1 acts on vectors. To make this more apparent, we let Xl + X2 = x, X2 + X3 = y, Xl + x3 = z, solve for Xl, x2, and X3 in terms of x, y, and z, andsubstitute in thepreceding equation to obtain rl(x,y,z) = !(x - y+z,x

+ y -z, -x + y+z).

Rewriting thisequation in terms of Xl, x2. and X3 gives T-

I(XI,X2,X3)

= !(XI -x2 +x3,XI +x2 -x3, -Xl +x2 +X3)'

We can easily verify thatT- 1T = 1 and thatTT- 1 = 1.

III

The following theorem, whose proof is left as an exercise, describes some properties of the inverse operator.

2.1.6. Theorem. The inverse of a linear operator is unique. 1fT and S are two invertible linear operators, then TS is also invertible, and

An endomorphism T : V ---> V is invertible onto another basis ofV.

if and only if it carries a basis of V

2.1.1 Polynomials of Operators With products and sums of operators defined, we can construct polynomials of operators. We define powers of T inductively as Tm TT"'-l Tm-IT for all positive integers m ::: I. The consistency of this equation (for m = I) demands that orO = 1. It follows that polynomials such as

=

p(T) = "01

=

+ "IT + "2T2 + ... + "n T"

can be defined.

ODTO l'WnJRJ'ROiANES!

I

M. E. T. U. USRARY

52

2. OPERATOR ALGEBRA

2.1.7. Example. Let TO : IR2 --+ ]R2 be the linear operator that rotates vectors in the xy-plane through the angle 0, that is, Te(x, y) = (x cos s - y sin e', x sln s

+ y cos s).

We are interested in powers of Te: y'

x'

T~(X, y) = Te(x cos s - y sin e', x stn s

+ y ccs s)

- y' sine, x' sine + y' cos B) «x cos s - y sin s) cos s - (x sinf + y coae) sine, (x cos s - y sinO) sin s + (x sinO + ycos sj cos s) (x cos 20 - y sin 20, x sin 20 + y cos20).

= (x' cos s =

=

Thus, T 2 rotates (x, y) by 20. Similarly, one can show that T~(X, y) = (x cos 30 - y sin 30, x sin 30 + ycos30),

and in general, l"9(x, y) = (x cosnB - y sinnB, x sinnB + y cos n(J). whichshowsthat 19 is a rotation of (x, y) through the angle ntJ, thatis, TO = TnO. This resultcould havebeen guessedbecause TO is equivalent to rotating (x, y) n times,eachtime by ananglee. III Negative powers of an invertible linear operator T are defined by T- m (T-I)m. The exponents of T satisfy the usual rules. In particular, for any two integers m and n (positive or negative), TmTn = ~+n and ~)n = r nn . The first relation implies that the inverse of T m is T- m . One can further generalize the exponent to include fractions and ultimately all real numbers; but we ueed to wait uutil Chapter 4, in which we discuss the spectral decomposition theorem.

2.1.8. Example. Let us evaluate Tin for the operator of theprevious example. First, let us find Tel (see Figure 2.1). We are looking for an operator sucb that Tel Te (x, y) or Tel (x cos s - y sin e', x sinO + y cos s) = (x, y). We define x' = x cos () -

= (x, y), (2.1)

y sin e and y' = x sin e + y cos e and solve x and y in terms of

x' and y' to obtain x = x' cos e + y' sin (J and y = -x' sin e +./ cos e. Substituting for x and y in Equation (2.1) yields T01(X',

y') = (x' cose + y' sine, -x' sine

+ y' cos e).

Comparing this with the actionof T8 in the previous example, we discoverthat the only difference betweenthe two operators is the sign of the sin e term, We conclude that TO 1 has the sameeffect as T-8' So we have and

T 8-n = (T-I)n 8 = (T-8 )n =T_n8·

2.1 ALGEBRA OF L(V)

53

"J8Cx,y)

(x,y)

Figure 2.1

The operator T(;I andits inverse as they act on a point in theplane.

Tin T9 = 1:

It is instructive to verifythat

y'

x'

TinT'J(x, y)

=

T(in(x cosn(} - y sin nO', x sinne + y cosn(})

=

(x' cosn8

=

+ y' sinnB, -x' sinnB + y' cos nO) «x cosnB - y sinnB) cosnf:J + (x sinn() + y cosn(}) sinne, - (x cos nO - y sin nO) sin nrJ + (x sin n() + y cos nfJ) cosnB)

= (x (cos2 nO + sin2 nO), y(sin 2 nO + cos2 nO)) = (x, Similarly, we can showthat T,JTiJn(x, y) = (x, y).

y).

l1li

One has to keep in mind that p (T) is not, in general, invertible, even if T is. In fact, the snm of two invertible operators is not necessarily invertible. For example, although T and -Tare invertible, their sum, the zero operator, is not.

2.1.2 Functions of Operators We can go one step beyond polynomials of operators and, via Taylor expansion, define functions of them. Consider an ordinary function f (x), which has the Taylor

expansion f(x) =

L 00

, k.

(x - xO)k d

k=Q

k

f I

-k

dx

.

x=xo

in which xo is a point where f (x) and all its derivatives are defined. To this function, there corresponds a fnnction of the operator T, defined as

(T- x ol )k

f(T)=fdk{1 k=O

dx

x=xo

k!

(2.2)

54

2. OPERATOR ALGEBRA

Becausethisseriesis aninfinite sumof operators, difficulties mayarise concerning its convergence. However, as will be shown in Chapter 4, f (T) is always defined for finite-dimensional vector spaces. In fact, it is always a polynomial in T. For the time being, we shall think of f (T) as a formal infinite series. A simplification results when the function can be expanded about x = O. In this case we obtain

L 00

f(T) =

k~O

dkfl d k x

x~O

Tk -k' .

(2.3)

.

A widely used function is the exponential, whose expansion is easily found to be 00

T

e

'"

exp(T) =

Tk

L-' k~O k!

(2.4)

2.1.9. Example. Let us evaluate exp(aT) wheu T : ][l.2 -+ ][l.2 is given by T(x,y) = (-y,x).

We can find a general formula for the action ofTn on (x, y). Start with n T2(x, y)

= 2:

= T(-y, x) = (-x, -y) = -(x, y) = -l(x, y).

Thus, T2 = -1. From T andT2 we can easily obtain higher powers of T. Forexample: T3 = T(T2) = -T, 'f'I = T2T2 = 1, and in general, T2n = (-I)nl

T2n+ 1 = (-I)nT

= 0, 1,2,

.

forn =0, 1,2,

.

for n

Thus, (aT)n

exp(aT) =

(aT)n

(aT)2k+l

00

00

(aT)2k

L -,- + nLeven -,= L , + k~O L --, n odd n. n. k=O (2k + I). (2k). 00

=L k=O

a2k+lT2k+l

00

=TL k~O

00

a2kT2k

00

, +L--, =L (2k + I). k~O (2k). k=O

(_I)k a2k+l (2k

+

, I).

00

+lL k~O

(_lla2k+1 (2k

+

00

I T+L I). k~O

(_I)k a2k

(_I)k a2k

"

(2k).

Thetwo series arerecognized as sina andcos a. respectively. Therefore, we get eaT

= Tsina + 1cosa,

whichshowsthat eaT is a polynomial (of first degree) in T. The action of eaT on (x, y) is given by e·T (x, y)

= (sinaT + cos ut) (x, y) = sin aT(x, y) + cos «t (x, y) = (sina)(-y, x) + (cosa)(x, y) = (-y sin a, x sin a) + (x cos a, ycosa) = (x cosa - y sin a, x sin a + y cosa).

,

(2k).

1

2.1 ALGEBRAOFL(V)

55

The reader will recognize the final expression as a rotation in the xy-plane through an angle a. Thus, we can think. of eaT as a rotation operator of angle a about the z-axis. In this context T is called the generator of the rotation. II

2.1.3 Commutators The result of multiplication of two operators depends on the order in which the operators appear. This means thatifT, U E .G(V),then TU E .G(V) and UT E .G(V); however, in general UT i= TU. When this is the case, we say that U and T do not commute. The extent to which two operators fail to commute is given in the following definition. commutator defined

2.1.10. Definition. The commutator [U, T] of the two operators Uand T in .G(V) is another operator in .G (V), defined as [U, T] es UT-TU.

An immediate consequence of this definition is the following: 2.1.11. Proposition. For S, T, U E .G(V) and a, fJ E IC (or ffi.), we have [U, T] = -[T, U],

antisymmetry

no

[aU, fJn = afJ[U, [S, T+ U] = [S, T] + [S, U],

linearity

[S+T, U] = [S, U] + [T, U],

linearity in the left entry

[ST, U] = S[T, U] + [S, U]T,

right derivation property

[S, TU] = [S, T]U + T[S, U], [[S, T], U] + [[U, S],

n + [[T, U], S] = O.

linearity in the right entry

left derivation property Jacobi identity

Proof In almost all cases the proof follows immediately from the definition. The only minor exceptions are the derivation properties. We prove the left derivation property: [S, TU] = S(TU) - (TU)S = STU - TUS + TSU - TSU '-v-'

=0

= (ST - TS)U + T(SU - US) = [S, T]U + T[S, U].

The right derivation property is proved in exactly the same way. A useful consequence of the definition and Proposition 2.1.11 is for m=O,±I,±2, .... In particular, [A, 1] = 0 and [A, A-I] = O.

D

56

2. OPERATOR ALGEBRA

2.2 Derivatives of Functions of Operators Up to this point we have heen discussing the algebraic properties of operators, static objects that obey certain algebraic rules and fulfill the static needs of some applications. However, physical quantities are dynamic, and if we want operators to represent physical quantities, we must allow them to change with time. This dynamism is best illustrated in quantum mechanics, where physical observables are represented by operators. Let us consider a mapping H : ffi. ---> ,c(V), which/ takes in a real number and gives out a linear operator on the vector space V. We denote the image of t E ffi. by H(t), which acts on the underlying vector space V. The physical meaning of this is that as t (usually time) varies, its image H(t) also varies. Therefore, for different atime-dependent values of t, we have different operators. In particular, [H(t), H(t')] "" afor t "" t', operator does not A concrete example is an operator that is a linear combination of the operators D commute with itself and T introduced in Example 1.3.4, with time-dependent scalars. To be specific, atdifferent times let H(t) = Dcos wt +Tsinwt, where w is a constant. As time passes, H(t) changes its identity from D to T and back to D. Most of the time it has a hybrid identity! Since Dand T do not commute, values of H(t) for differeuttimes do not necessarily commute. Of particular interest are operators that can be written as exp H(t), where H(t) is a "simple" operator; i.e., the dependence of H(t) on t is simpler than the corresponding dependence of exp H(t). We have already encountered such a situation in Example 2.1.9, where it was shown that the operation of rotation around the z-axis could be written as expaT, and the action ofT on (x, y) was a great deal simpler than the corresponding action of exp aT. Such a state of affairs is very common in physics. In fact, it can be shown that many operators of physical interest can be written as a product of simpler operators, each being of the form exp aT. For example, we know from Euler's theorem in mechanics that an arbitrary rotation in three dimensions can be written as a product of three simpler rotations, each being a rotation through a so-called Euler angle about an axis. derivative of an operator

2.2.1. Definition. For the mapping H : ffi. ---> ,(,(V), we define the derivative as dH = lim H(t dt ""-+0

+ M) -

H(t).

f!,.t

This derivative also belongs to ,c(V).

As long as we keep track of the order, practically all the rules of differentiation apply to operators. For example, d dU dT dt (UT) = dt T + U dt . 2Strictly speaking, the domain ofH must be an interval [a, b] of the real line. because Hmay not be definedfor all R. However, for our purposes, such a fine distinction is Dot necessary.

2.2 DERIVATIVES DF FUNCTIONS OF OPERATORS

57

We are not allowed to change the order of multiplication on the RHS, not even when both operators being multiplied are the same on the LHS. For instance, if we let U = T = Hin the preceding equation, we obtain d

2

dH

dH

dt

dt

-(H)= -H+Hdt

This is not, in general, equal to 2H!fJ/-. 2.2.2. Example. Letus findthederivative of exp(tH),whereHis independentof t. Using Definition 2.2.1, we have

d -exp(tH) dt

lim exp[(t + Ll.t)Hj = .6.t-+0 IJ..t

exp(tH)

.

However,for infinitesimal I::1t we have

cxpltr + Ll.t)H] - exp(tH) = etHel!.tH _ et H = et H(1 + HLl.t) _ etH = etHHLl.t.

Therefore, d

etHHLl.t

- exp(tH) = lim - - -

dt

Ll.t

l!.HO

= etHH.

Since Hand et H commute.e we also have d

- exp(tH) = Het H. dt

Note that in deriving the equation for the derivative of et H , we have used the relation et He..6..tH = e(t+Lit)H. This may seem trivial.but it will be shown later that in general, ~#~~.



Now let us evaluate the derivative of a more general time-dependent operator, exp[H(t)]: d

.

- exp[H(t)] = Ion

dt

exp[H(t

+ Ll.t)] Ll.t

l!.HO

exp[H(t)]

.

If H(t) possesses a derivative, we have, to the first order in Ll.t, H(t

+ Ll.t)

= H(t)

d

+ Ll.t dt H,

and we can write exp[H(t + Ll.t)] = exp[H(t) + Ll.tdH/dt]. It is very tempting to factor out the exp[H(t)] and expand the remaining part. However, as we will see presently, this is not possible in general. As preparation, consider the following example, which concerns the integration of an operator.

58

2. OPERATOR ALGEBRA

evolution operator

2.2.3. Example. The Schrtidinger equation i ;t It(t» = H It(t» can be turned into an operator differential equation as follows. Define the so-called evolution operator U(t) by It(t» = U(t) It(O», and substitute in the Schrtidinger eqnation to obtain i ~U(t)

at

It(O»

= HU(t)

It(O» .

Ignoring the arbitraryvector [t(O» results in a differential equation in U(t). For the purposes ofthis example.Iet usconsider anoperator differential equation of the form dU / dt = HU(t), where H is not dependent on t, We can find a solution to such an equation by repeated differentiation followedby Taylor seriesexpansion. Thus, d 2U d ~ = H- U = H[HU(t)] = H2U(t), dt dt d 3U

d

d

dt

dt

dt

- 3 = -[H 2U(t)] = H2_U = H3U(t).

In general dnU/dt n = HnU(t). Assuming that U(t) is well-defiued at t = 0, the above relations say that all derivatives of U(t) are also well-defined at t = O. Therefore, we can expand U(t) arouud t = 0 to obtain U(t) =

=

tn (dnU) tn L, -d n = L ,HnU(O) n=O n. t t=O n=O n. 00

00

(~

U:r)

U(O) = e'HU(O).

III

Let us see under what conditions we have exp(5 + T) = exp(5) exp(T). We consider only the case where the commutator of the two operators commutes with both of them: [T, [5, Tj] = 0 = [5, [5, T]].NowconsidertheoperatorU(t) = e'Se'Te-'(SH) and differentiate it using the result ofExarnple 2.2.2 and the product rule for differentiation:

~U =

5e'setTe-t(SH)

dt

+ etSTe'Te-'(S+T)

_ e'Se'T (5

+ T)e-t(SH)

= 5e'Se'T e-t(SH) _ etSetT5e-t(SH).

(2.5)

The three factors of U(t) are present in all terms; however. they are not always next to one another. We can switch the operators if we introduce a commutator. For instance, e'T5 = 5e tT + letT, 5]. Itis leflas aproblemforthereaderto show that if [5, Tj commutes with 5 and T, then letT, 5] = -t[5, T]etT, and therefore, etT5 = 5e,r - t[5, T]etT. Substituting this in Equation (2.5) and noting that e'55 = SetS yields dU/dt = t[5. TjU(t). The solution to this equation is

U(t)

= exp

c;

[5, T])

=>

e'Se'Te-t(SH)

= exp

c;

[5, T])

3Thisis a consequence of a moregeneral result that if two operators commute, anypairof functions of those operators also commute (see Problem 2.14).

2.2 DERIVATIVES OF FUNCTIONS OF OPERATORS

59

because U(O) = 1. We thus have the following: 2.2.4. Proposition. Let S, T E £'(V). If [S, [S, T]] Baker-CampbellHausdorff formula

= 0 = [T, [S, TJ], then the

Baker-Campbell-Hausdorffformula holds: (2.6) In particular, e'Se'T = e,(s+n ifand only if[S,

TJ = o.

If t = I, Equation (2.6) reduces to eSeTe-(l/2)[S,T] = eS+T.

(2.7)

Now assume that both H(t) and its derivative commute with [H, dH/dt]. Letting S H(t) and T f!J.tdH/dt in (2.7), we obtain

=

eH(t+.6.t)

=

= e H(t)+.6.tdH/dt = eH(I) elll(dH/dl)e-[H(I),lltdH/dl]/2.

For infinitesimal f!J.t, this yields H(I+IlI) e =eH(I)

=e

(1 +f!J.t~~) (1- ~f!J.t[H(t), ~~])

H(I){l

+f!J.t~~ - ~f!J.t[H(t), ~~]},

and we have

~eH(I)

dt

= eHd H _ !eH dt 2

[H ' dH]. dt

We can also write eH(t+IlI)

= e[H(I)+lltdH/dl] = e[llldH/dl+H(I)] = e[llldH/dl]eH(t)e-[lltdH/dl,H(ll]/2,

which yields

~eH(ll

dt

= dH eH + !eH dt 2

[H ' dH]. dt

Adding the above two expressions and dividing by 2 yields the following symmetric

expression for the derivative:

~eH(ll

dt

antlcommutator

=! 2

(dH eH +eHdH) =! {dH eH} dt dt 2 dt ' ,

where IS, T} sa ST + TS is called the anticommutator of the operators S and T. We, therefore, have the following proposition.

60

2. OPERATOR ALGEBRA

2.2.5. Prepositlon, Let H : R --> .G (V) and assume that H and its derivative commute with [H, dH/dt]. Then

=!2 {dH eH} dt ' .

!£eH(t) dt

In particular, if[H, dH/dt] = 0, then

= dH eH = eH dH .

!£eH(t) dt

dt

dt

A frequently eucountered operator is F(t) = etABe- tA, where A and B are t-independent. It is straightforward to show that dF

-

dt

= [A, F(t)]

d -[A, F(t)] = [ A, dF] . dt dt

and

Using these results, we can write d 2F -2 dt

d = dt -[A, F(t)] = [A, [A, F(t)]] sa A2[F(t)],

and in general, dnF/dt n = An[F(t)], where An[F(t)] is defined inductively as An[F(t)] = [A, An-1[F(t)]], with AO[F(t)] '" F(t). For example, A3[F(t)]

= [A, A2[F(t)]] = [A, [A, A[F(t)]]] = [A, [A, [A, F(t)]]].

Evaluating F(t) and all its derivatives at t = 0 and substituting in the Taylor expansion about t = 0, we get

=L 00

F(t)

n=O

dRFI ,n.tn -.dt

tn ,An [F(O)] n=O n.

=L 00

1=0

t" ,An[B]. n=O n.

=L 00

That is, etABe- tA =

tn L _An[B] ea B + t[A, B] + -[A, [A, B]] + .... n~O n! 2! t2

00

Sometimes tltis is written symbolically as

where the RHS is merely an abbreviation of the infinite sum in the middle. For t = I we obtain a widely used formula: eABe- A = eA[B] =

I ) I L ,An [B] '" B + [A, B] + ,[A, [A, B]] + .... n. 2. 00

(

n=O

61

2.3 CONJUGATION OF OPERATORS

If A commutes with [A, BJ, then the infinite series truncates at the second term, and we have etABe- tA = B + t[A, BJ. For instance, if A and B are replaced by D and T of Example 1.3.4, we get (see Problem 2.3)

etOTe-tO = T + t[D, T] = T + t1. generator of translation

momentum as generator of translation

The RHS shows that the operator T has been translated by an amount t (more precisely, by t times the unit operator). We therefore call exp(tD) the translation operator of T by t, and we call D the generator of translation. With a little modification T and D become, respectively, the position and momentum operators in quantum mechanics. Thus,

2.2.6. Box. Momentum is the generator oftranslation in quantum mechan-

ics. But more of this later!

2.3 Conjugation of Operators We have discussed the notion of the dual of a vector in conjunction with inner products. We now incorporate linear operators into this notion. Let Ib) , [c) E '\7 and assume that [c) = T Ib). We know that there are linear functionals in the dual space '\7* that are associated with (Ib))t = (bl and (Ic))t = (c], Is there a linear operator belonging to £'('\7*) that somehow corresponds to T? In other words, can we find a linear operator that relates (bl and (cl just as T relates Ib) and [c}? The answer comes in the following definition. adjoint ofan operator

2.3.1. Definition. Let T E £'('\7) and 10) ,Ib} E '\7. The adjoint, or hermitian conjugate, ofT is denoted by rt and defined by (01 T Ib)* = (bl Tt 10) .

(2.8)

The LHS of Equation (2.8) can be written as (01 c)* or (c] a), in which case we can identify (2.9) This equation is sometimes used as the definition of the hermitian conjugate. From Equation (2.8), the reader may easily verify that 1t = 1. Thus, using the unit operatorfor T, (2.9) justifies Equation (1.12). Some of the properties of conjugation are listed in the following theorem, whose proof is left as an exercise.

62

2. OPERATOR ALGEBRA

2.3.2. Theorem. Let U, T E .(,(V) and a E C. Then 1. (U + T)t = Ut 3. (OIT)t =OI*T t.

+ Tt.

2. (UT)t = TtU t. 4. «T)t)t = T.

The last identity holds for finite-dimensional vector spaces; it does not apply to

infinite-dimensional vectorspaces in general. In previous examples dealing with linear operators T : Ill" -> Ill", an element of Ill" was denoted by a row vector, such as (x, y) for Ill 2 and (x, y, z) for Ill3. There wasno confusion, because we wereoperating only in V. However, sinceelements of both V and V* are required when discussing T, T*, and rt, it is helpful to make a distinction between them. We therefore resort to the convention introduced in Example 1.2.3 by which .

2.3.3. Box. Ketsare representedas column vectors and bras as row vectors.

2.3.4. Example. Let us find the hermitianconjugateof the operatorT : C3 -> C3 given by T

(~~) =

(at i::~~ a

{t3

3

) .

al - a2

+ ia3

Introduce

la}

=

GD

and

with dual vectors (al = (aj Equation(2.8) to findTt:

Ib) =

(~D

a2 a3)and (bl = (13j

132 133), respectively. We use

2.4 HERMITIAN AND UNITARY OPERATORS

63

Therefore, we obtain

III

2.4 Hermitian and Unitary Operators The process of conjugation of linear operators looks much like conjugation of complex numbers. Equation (2.8) alludes to this fact, and Theorem 2.3.2 provides further evidence. It is therefore natural to look for operators that are counterparts of real numbers. One can define complex conjugation for operators and thereby construct real operators. However, these real operators will not be interesting because-as it turns out-they completely ignore the complex character of the vector space. The following altemative definition makes use of hermitian conjugation, and the result will have much wider application than is allowed by a mere complex conjugation. hermitian and antl-hermltlan

operators

2.4.1. Definition. A linear operator H E L(V) is calledhermitian, or self-adjoint, ifHt = H. Similarly, A E L(V) is called anti-hermitian if At = -A

Charles Hermite (1822-1901), one of the most eminent French mathematicians of the nineteenth century, was particularly distinguished for the cleanelegance and high artistic quality of his work. As a student, he courted disaster

by neglecting his routine assigned workto studythe classic masters of mathematics; andthough he nearly failedhis examinations, he became a first-rate creative mathematician whilestill in his earlytwenties. In 1870he was appointed to a professorship at the Sorbonne, where he trained a whole generation of well-known French mathematicians, including Picard, Borel, and Poincare. Thecharacter of his mindis suggestedby aremark ofPoincare: "Talk withM. Hermite. Heneverevokesaconcrete image,yet yousoonperceivethatthemostabstract entities areto him likelivingcreatures." Hedislikedgeometry, butwasstrongly attracted to number theory andanalysis,andhis favorite subject was elliptic functions, where these two fields touch in many remarkable ways. Earlier in the century the Norwegian genius Abel had proved that the general equation of the fifth degreecannotbe solved by functions involvingonly rational operations androotextractions. Oneof Hermite's mostsurprising achievements (in 1858) was to showthat thisequation can be solved by ellipticfunctions. His 1873 proof of the transcendence of e was another high point of his career," If he hadbeenwilling to dig even deeper intothis vein, he couldprobably havedisposedof tt as 4Transcendental numbers arethose that arenot roots of polynomials withinteger coefficients.

64

2. OPERATOR ALGEBRA

well,but apparently he had hadenoughof a goodthing.As he wroteto a friend, "I sballrisk nothing on anattempt to provethetranscendence of the number n, If others undertake this enterprise, no one will be happier than I at their success, butbelieveme, my dear friend, it will not fail to cost themsome efforts." As it turned out, Lindemann's proof nine yearslater restedon extendingHermite's method. Several of his purely mathematical discoveries hadunexpected applications many years later to mathematical physics. For example, the Hermitian forms and matrices that he invented in connection with certain problems of number theoryturned out to be crucial for Heisenberg's 1925formulation of quantum mechanics, andHermite polynomials (seeChapter 7) areuseful in solving Schrodinger's wave equation.

The following observations strengthen the above conjecture that conjugation of complex numbers and hermitian conjugation of operators are somehow related.

expectation value

2.4.2. Definition. The expectation value (T)a of an operator T in the "state" la) is a comp/ex number defined by (T)a = (aITla) .. The complex conjugate of the expectation value is5 (T)* = (al T la)*

= (al rt la).

In words, r', the hermitian conjugate of T, has an expectation value that is the complex conjugate of the latter's expectation value. In particular, ifT is hermitianis equal to its hermitian conjugate-its expectation value will be real. What is the analogue of the known fact that a complex number is the sum of a real number and a pure imaginary one? The decomposition

shows that any operator can be written as a sum of a hermitian operator H = ~ (T + Tt) and an anti-hermitian operator A = ~ (T - Tt). We can go even further, because any anti-hermitian operator A can be written as A = i(-iA) in which -iA is hermitian: (-iA)t = (-i)*At = i(-A) = -iA. Denoting -iA by H', we write T = H + iH', where both H and H' are hermitian. This is the analogue of the decomposition z = x + i Y in which both x and y are real. Clearly, we should expect some departures from a perfect correspondence. This is due to a lack of conunutativity among operators. For instance, although the product of two real numbers is real, the product of two hermitian operators is not, in general, hermitian:

5Whenno risk of confusionexists, it is common to drop the subscript "a" andwrite (T) for the expectation valueofT.

2.4 HERMITIAN AND UNITARY OPERATORS

65

We have seen the relation between expectation values and conjugatiou properties of operators. The following theorem completely characterizes hermitian operators in terms of their expectation values: 2.4.3. Theorem. A linear transformation H on a complex inner product space is hermitian if and only if (al H la) is real for allla).

Proof We have already pointed out that a hermitian operator has real expectation values. Conversely, assume that (al H la) is real for all ]e). Then (al H la)

= (al H la)* = (al Ht la)

(al H - Ht la)

{}

By Theorem 2.1.4 we must have H - Ht =

=0

Via) .

o.

D

2.4.4. Example. In thisexample, we illustrate theresult of theabove theorem with2 x 2 matrices. The matrix H = C? is hermitian'' andacts on «:2. Let us take an arbitrary

c/)

vector la)

= (~P and evaluate (al H la). We have

Therefore, (al H la) = (ai = iaial

a

2) = -iaia2 + iazat z)(-.ia ,a!

+ (ia20:1)* =

2 Re(io:iat),

and (al H la) is real.

Forthemostgeneral 2 x 2 hermitian matrix H = have Hla) =

(;* ~), where

a

and y arereal, we

(;*

and (al H la) = (Oli

az)

(pOl*OI! ++P(2) = ai(aa! at

ya2

+ P0l2) + OIZ(P*OI! + y(2)

= ala!1 2 + aiP0I2 + aZp*a! + Y la21 2 = 0I1a!1 2 + YI0I212 + 2 Re(aip0I2). Again (al H la) is real.

positive operators

III

2.4.5. Definition. An operator A on an inner product space is called positive (written A:::: 0) if A is hermitian and (al Ala) :::: Ofor allla).

6Weassume thatthereader hasa casualfamiliarity withhermitianmatrices. Thinkof an n x n matrix as a linearoperator that actson columnvectorswhose elementsarecomponents of vectorsdefinedin the standard basis of en or R". A hermitian matrix thenbecomes a hermitian operator.

66

2. OPERATOR ALGEBRA

2.4.6. Example. Ali example of a positive operator is the squareof a hermitian operator.? We note that for any hermitian operator H andany vector [c), we have (c] H2 1a} = (al HtH la) = (Hal Ha) ~ obecauseof thepositivedefiniteness of the innerproduct. III

positive definite operators

An operator T satisfying the extra condition that (a IT la) = 0 implies la) = 0 is called positive definite. From the discussion of the example above, we conclude that the square of an invertible hermitian operator is positive definite. The reader may be familiar with two- and three-dimensional rigid rotations and the fact that they preserve distances and the scalar product. Can this be generalized to complex inner product spaces? Let jz) , Ib) E V, and let U be an operator on V that preserves the scalar product; that is, given Ib') = U Ib) and la') = U la), then (a'i b') = (al b). This yields

(a'l b')

= «al Ut)(U Ib»

= (al UtU Ib)

= (al b) = (aI1Ib).

Since this is truefor arbitrary la) and Ib), we obtain u'u = 1. In the next chapter, when we introduce the concept of the determinant of operators, we shall see that this relation implies that U and ut are both Inverrible," with each one being the inverse of the other. unitary operators

2.4.7. Definition. Let V be afinite-dimensional inner product space. An operator U is called a unitary operator ifU t = U- I. Unitary operators preserve the inner product o/V. 2.4.8. Example. The linear transformation T : C3 --> C 3 givenby T

I:~) =

\;;3

(

~li~2i'::~~:0./6

(al {al - a2 + a3 + i(al

)

+ cz + (3)}/./6

is unitary. In fact,let

and with dual vectors (al = (at ai a3') and (hi = (Pi Pi P3'), respectively. We use Equation(2.8) and the procedureof Example2.3.4 to find rt, The result is al

a2

v1+./6+ ial

ia2

v1-./6-

a3(1 - i)

./6

+ i) ./6

a3(l

~This is further evidencethathermitian operators areanalogues of realnumbers: The square of anyrealnumber is positive. This implication holds only for finite-dimensional vectorspaces.

2.5 PROJECTION OPERATORS

67

Thus TIt = 1. Similarly, we can show that TtT = 1 and therefore that T is unitary.

II

and we can verify that

2.5 Projection Operators We have already considered subspaces briefly. The significance of subspaces is that physics frequently takes place not inside the whole vector space, but in one of its subspaces. For instance, although motion generally takes place in a threedimensional space, it may restrict itself to a plane either because of constraints or due to the nature of the force responsible for the motion. An example is planetary motion, which is confined to a plane because the force of gravity is central. Furthermore, the example of projectile motion teaches us that it is very convenient to "project" the motion onto the horizontal and vertical axes and to study these projections separately. It is, therefore, appropriate to ask how we can go from a full space to one of its subspaces in the context of linear operators. Let us first consider a simple example. A point in the plane is designated by the coordinates (x, y). A subspace of the plane is the x-axis. Is there a linear operator.? say Px , that acts on such a point and somehow sends it into that subspace? Of course, there are many operators from ]Rz to R However, we are looking for a specific one. We want Px to project the point onto the x-axis. Such an operator has to act on (x, y) and produce (x, 0): Px(x, y) = (x, 0). Note that if the point already lies on the x-axis, Px does not change it. In particular, if we apply Px twice, we get the same result as if we apply it ouly once. And this is true for any point in the plane. Therefore, our operator must have the property P~ = Px. We can generaIize the above discussion in the following deflnition.I'' projection operators

2.5.1. Definition. A hermitian operator PEl:., (V) is calleda projectionoperator ifPZ = P. From this definition it immediately follows that the only projection operator with an inverse is the identity operator. (Show this!) Consider two projection operators PI and Pz. We want to investigate conditions under which PI + Pz becomes a projection operator. By definition, PI + Pz = (PI + Pz)z = + PI Pz + PZPI + P~. So PI + Pz is a projection operator if and ouly if

pi

(2.10) 9We want this operator to preserve the vector-space structure of the plane and the axis. l°lt is sometimes useful to relax the condition ofherrniticity. However, in this part of the book, we demand that P be hermitian.

68

2. OPERATOR ALGEBRA

MUltiply this ou the left by PI to get PfP2

+ PIP2PI = 0

=}

PIP2

+ PIP2PI = O.

Now multiply the same equatiou ou the right by PI to get PIP2PI

+ P2Pf = 0

=}

PIP2PI

+ P2PI = O.

These last two equatious yield (2.11) The solutiou to Equatious (2.10) aud (2.11) is PIP2 = P2PI = O. We therefore have the following result.

orthogonal projection operators

2,5.2. Proposition. Let PI, P2 E .G(V) be projection operators. Then PI + P2 is a projection operator if and only if PIP2 = P2PI = O. Projection operators satisfying this condition are called orthogonal projection operators. More geuerally, if there is a set {P;}i~1 ofprojectiou operators satisfyiug Pi { , J 0

p.p. -

ifi = j, ifi '" j,

theu P = I:7::1 Pi is also a projectiou operator. Giveu a uormal vector Ie), oue cau show easily that P = [e) (e] is a projectiou operator: • Pishermitiau: pt = (Ie) (el)t = «el)t(le»t = Ie) (e].

---

• P equals its square: p 2 = (le) (el)(le) (el) = Ie) (el e) (e] = lei (e], =1

ill fact, we cantake au orthouormal basis B = {lei) }~I and coustruct a set of projectiou operators {Pi = lei) (eil}~I' The operators Pi are mutually orthogoual. Thus, their sum I:~I Pi is also a projectiou operator. compieteness relation

2.5.3. Propositiou. Let B = {lei}}~1 be an orthonormal basis for VN. Then the set {Pi = lei) (eil}~1 consists of mutually orthogonal projection operators, and I:~I Pi = I:~I lei) (ei I = 1. This relation is called the completeness relation. Proof The mutual orthogouality of the Pi is au immediate cousequeuce of the orthouormality of the lei). To show the secoud part, consider au arbitrary vector la), writteu in terms of lei): la) = I:f=1 a j lej i. Apply Pi to both sides to obtain

2.5 PROJECTION OPERATORS

69

Therefore, we have

11a) = tai lei) =

tPi

la) = (tPi) la).

Since this holds for an arbitrary

la), the two operators must be equal.

D

If we choose only the first m < N vectors instead of the entire basis, then the projection operator p(m) '" I:~=I lei) (eil projects arbitrary vectors into the subspace spanned by the first m basis vectors (lei) li':"I' In other words, when p(m) acts on any vector la) E V, the result will be a linear combination of only the first m vectors. The simple proof of this fact is left as an exercise. These points are illustrated in the following example. 2.5.4. Example. Consider three orthonormal vectors

(lei) 1;=1 E lW,3 given by

Theprojection operators associated with each of these can be obtained by notingthat (ei I is a row vector.Therefore, PI

= leI) (eIl = ~

(D

0)

10)=2:I (I1 I1 0.

(I

.

0

0

0

Similarly,

pz=H~I)(l

2)=H~1 ~: ~2)

-I

and

I

I)=~ ( -I 3 -I

-I

I

-I) I I

I

.

Note thatPi projects onto the line along lei}. This can be testedby letting Pi act on an arbitrary vectorandshowing that the resulting vectoris perpendicular to the other two vectors. Forexample, let P2 act on anarbitrary columnvector: la)

==

Pz (;)

=

z

~ (~I ~I 2

-2

!2) (;)z 4

=

~ (:x-:/_2~z).

2x-2y+4z

We verify that]e} is perpendicular to both let) and [eg}:

I

(ella)= vrz(1

I

Y+2Z) I(X-x + Y - 2z = 2x-2y+4z

0) (;

O.

70

2. OPERATOR ALGEBRA

Similarly, (e31 a)

=

t

(2

O. So indeed, la) is atong le2). We can find the operator that projects onto the plane fanned by leI) and le2). This is

PI

+ P2 =

t

- I 3 I

2

-t

When this operator acts on an arbitrary column vector, it produces a vector lying in the plane of 1eJ) and le2), orperpeodicularto le3):

Ib) == (PI

+ P2) ( YX) = 3"t Z

(2

I t-I ) (X) Y

I I

2

-t

2

Z

= 3"t GX+Y+Z) X + 2y - Z • - Y + 2z

It is easy to show that (e31 b) = O.The operators that project onto the other two planes are

obtainedsimilarly. Finalty, we verifyeasily that PI

+ P2 + P3 = (~

o

~ ~~) = 1.

11II

0

2.6 Operators in Numerical Analysis In numerical calculations, limitiog operations involving infinities and zeros are replaced with finite values. The most natural setting for the discussion of such operations is the ideas developed in this chapter. In this section, we shall assume that all operators are invertible, and (rather sloppily) manipulate them with no mathematical justification.

2.6.1 Finite-Difference Operators

forward, backward, and central difference operators

In all numerical manipulations, a function is considered as a table with two columns. The first column lists the (discrete) values of the independent variable Xi, and the second column lists the value of the function I at Xi. We often write Ii for I(xi). Three operators that are in use in numerical analysis are the forward difference operator ~,the backward difference operator V (not to be confused with the gradient), and the central difference operator 6. These are defined as follows: ~Ii sa 11+1 -

fi,

Vfi

es

Ii -fi-I, (2.12)

The last equation has only theoretical significance, because a half-step is not used in the tabulation of functions or in computer calculations. Typically, the data are equally spaced, so Xi+1 - Xi h is the same for all i, Then Ii±1 I(xi ± h), and we define Ii±lj2 sa I(xi ± hI2).

=

=

2.6 OPERATORS IN NUMERICAL ANALYSIS

71

We can define products (composition) of the three operators. In particular, A 2 is given by

A 2 f;

= A(f1+1 -

ft)

= 11+2 -

2/1+1

+ f;.

(2.13)

Similarly,

V 2 f; = V(f; - li-l) = Ii - 2li-l + Ii-a.

tP f;

= f;+J - 21i

+ Ii-I. (2.14)

We note that

tP f;

shifting and averaging operators

= f;+1 - f; - (f; - Ii-I) = (A - V)li '---v-'

'-,,-'

=Afi

=V!i

=}

(j2 = A-V.

This shows that the three operators are related. It is convenient to introduce the shifting and averaging operators, respectively E and p" as

E/(x) = I(x

+ h),

p,1(x) =

~ [I (x +~) +I

Note that for any positive integer n, EnI(x) = I(x any real number a: a

E

(x -

~)].

(2.15)

+ nh). We generalize this to

I(x) = I(x + ah).

(2.16)

All the other finite-difference operators can be written in terms of E:

A=E-1,

V=1-E- 1,

(j=E I/ 2_E-1/2,

p,=~(El/2+E-l/2)

(2.17) The first two equations of (2.17) can be rewritten as

E=l +A,

E = (1 - V)-l

(2.18)

We can obtain a useful formula for the shifting operator when it acts on polynomials of degree n or less. First note that

1 - V n +1 = (1 - V)(l + V + ... + V n ) . But V n +1 armihilates all polynomials of degree n or less (see Problem 2.33). Therefore, for such polynomials, we have

1 = (1 - V)(l + V + ., . + Vn ) ,

72

2. OPERATOR ALGEBRA

which shows that E obtain

= (1 -

= 1 + V + ... + Vn . Now let n -+

V)-I

00

and

00

E = (1 - V)-I =

LV

k

(2.19)

k~O

for polynomials of any degree and-hy Taylor expansion-for any (well-behaved) function. 2.6.1. Example. Numerical interpolation illustrates the use of the formulas derived above. Suppose that we are given a table of the values of a function, and we want the valueof thefunction foran x locatedbetween two entries. Wecannot use thetable directly, butwe mayuse thefollowing procedure. Assumethat thevalues of the function f are given for Xl, X2, ... , xi, ... , and we are interested in the valueof the function for x suchthat xi < x < xi+ 1. This corresponds to Ii+r. where0 < r < 1. Wehave r

r

li+r=Eli=(1+~)/i=

(

r(r - I)

1+r~+--2-~

2

+ ... )

j;.

(2.20)

In practice, the infinite sumis truncated after a finite number of terms. If only two tenus arekept, we have

li+r "" (1 + r~)li = Ii

+ r(fi+1

- Ii) = (I - r)/i

+ rli+l.

(2.21)

t.

In particular. for r = Equatiou (2.21) yields li+l/2 "" t(fi + li+I). which states the reasonable result that thevalueatthemidpoint is approximately equalto the average of the values at the endpoints. If the third term of the series in Equation (2.20) is also retained. theu

r(r - I) 2 r(r - I) 2] li+r"" [ 1 +r~+ --2-~ Ii = Ii +r~/i + --2-~ Ii . r(r - I) = Ii + r(fi+l - /i) + --2-(1i+2 - 2li+l + Ii) =

For r

~-~(I-~

2

Ii

+ r(2 -

r)li+1

(2.22)

r~-I)

+ --2-11+2'

= t. that is. at the midpoint betweeu Xi and Xi +1. Equation (2.22) yields

1i+1/2 "" ~Ii

+ lli+l -

!/i+2.

whichturns outto he abetter approximation than Equation (2.21). However, it involves not onlythetwopoints on either sideof x butalsoa relatively distant point, Xi +2, If we wereto retain terms up to ~k for k > 2. theu Ii+r would be giveu in terms of h. Ii+j. ...• li+k> and the result would be more accurate than (2.22). Thus, the more information we have about the behavior of a function at distant points; the better we can approximate it at x E (Xj,Xi+l). Theforegoing analysis was basedon forward interpolation. Wemaywantto use backwardinterpolation, where fi-r is sought for 0 < r < 1.In such a case we use thebackward difference operator

Ii-r

= (E- 1)rIi = (1 -

V)r Ii

- I) 2 ) = (1 -rrV(+r --2-V +... h-

III

2.6 OPERATORS IN NUMERICAL ANALYSIS

73

2.6.2. Example. Let us check the conclusion made in Example 2.6.1 using a calculator and a specific function, say sinx. A calculator gives sin(O.I) = 0.0998334, sin(O.2) = 0.1986693, sin(0.3) = 0.2955202. Suppose that we wantto find sin(O.l5) by interpolation. Using Equation (2.21) with r = we obtain

i,

sin(0.15) "" ![sin(O.I) + sin(0.2)] = 0.1492514. On the other hand, using (2.22) with r =

sin(0.15) ""

i yields

i sin(O.I) + i sin(0.2) - ksin(0.3) = 0.1494995.

The value of sin(O.l5) obtainedusing a calculator is 0.1494381. It is clear that (2.22)gives a better estimate than (2.21). III

2.6.2 differentiation and integration operators

Differentiation and Integration Operators

The two most important operations of mathematical physics can be writtett itt temns of the finite-difference operators. Define the differetttiatiott operator 0 and the mtegration operator J by

Of(x)

==

f'(x),

l

Jf(x) es x

x +h

f(t) dt.

(2.23)

Assuming that 0- 1 exists, we note that f(x) = 0- 1[f'(x)]. This shows that 0- 1 is the operatiott of antidifferentiation: 0- 1 f(x) = F(x), where F is any primitive of f. Ott the other hand, AF(x) = F(x +h) - F(x) = Jf(x). These two equations and the fact that J and 0 commute (reader, verify!) show that

Ao- 1 = J

=}

JO = OJ = A = E -1.

(2.24)

Using the Taylor expansion, we can write

"exact" relation between shiffing and differentiation operators

(2.26) 2.6.3. Example. Let us calculate cos(O.I), considering cosx to be (dldx)(sinx) and usingthe valuesgiven in Example 2.6.2. Using Equation (2.26) to secondorder, we get

74

2. OPERATOR ALGEBRA

Thisgives cos(O.I) '" 1[-0.295520 + 4(0.198669) - 3(0.099833)] = 0.998284.

Incomparison. thevalue obtained directly from a calculator is 0.995004. "exact" relation between integration and difference operators

iii

The operator J is a special case of a more general operator defined by Jaf(x)

=

t: x

f(t) dt

= F(x + ah) -

F(x)

= (Ea -l)F(x) = (Ea _1)D- 1f(x),

or J

= (Ea a

-1)D- 1

Ea - 1

(1+a)a-1

= hIn-E- = h-'-c:--,--'---:-,-1n(1 + a) ,

(2.27)

where we nsed Equation (2.26).

2.6.3 Numerical Integration

J:

Suppose that we are interested in the numerical value of f(x) dx. Let Xo sa a and XN == b, and divide the interval [a, b] into N equal parts, each of length h = (b - a)jN. The method commonly used in calcnlating integrals is to find the integral f(x) dx, where a is a suitable number, and then add all such integrals. More specifically,

J:,;+ah

1=

t: t: f(x)dx+

f(x)dx+ ... +

xo+ah

XQ

r:

where M is a suitably chosen number. In fact, since x N = Xo M a = N. We next employ Equation (2.27) to get 1= Jafo

+ Jafa + ...+ Jaf(M-l)a =

f(x)dx,

xo+(M-l)ah

Ja

+ N h,

(~l fka) ,

we have

(2.28)

k~O

where fka sa f(xo + kah). We thus need an expression for Ja to evaluate the integral. Such an expression can be derived elegantly, by noting that

a

E - 1 I = -E la = - = -J a, DinED InE h a

1

S

S

E ds

[by Equation (2.27)]

so that

(2.29)

2.6 OPERATORS IN NUMERICAL ANALYSIS

75

where we expanded (1 + A Y using the binomial infinite series. Equations (2.28) and (2.29) give the desired evaluation of the integral. Let us make a few remarks before developing any commonly used rules of approximation. First, once h is set, the function can be evaluated only at Xo + nh, where n is a positive integer. This means that fn is given only for positive iotegers n. Thus, in the sum in (2.28) ka must be an integer. Since k is an integer, we conclude that a must be an integer. Second, since N = Ma for some integer M, we must choose N to be a multiple of a. Third, if we are to be able to evaluate Jaf(M-l)a [the last term io (2.28)], J a cannot have powers of A higher than a, because An f(M-l)a contaics a term of the form f(xo

trapezoidal rule for numerical integration

+ (M -

l)ah

+ nh) =

f(XN

+ (n -

a)h),

which for n > a gives f at a point beyond the upper limit. Thus, in the power-series expansion of J a, we must make sure that no power of A beyond a is retained. There are several specific Ja's commonly used in numerical integration. We will consider these next. The trapezoidal rnle sets a = 1. Accordiog to the remarks above, we therefore retain terms up to the first power in the expansion of J a. Then (2.29) gives JI = J = h(1 + ~A). Substituting this in Equation (2.28), we obtain N- l

1= h(1

+ ~A) ( ~ fk

h N-l

)

N-l

=h

~[!k + ~(fk+l

- fk)]

h

="2 L(fk + fk+l) = "2(fo+2fl + .. ·+2fN-l +fN). (2.30)

k~O

Simpson's one-third rule for numerical integration

Simpson's one-third rule sets a = 2. Thus, we have to retain all terms up to the A 2 term. However, for a = 2, the third power of A disappears in Equation (2.29), and we get an extra ''power'' of accuracy for free! Because of this, Simpson's one-thirdrule is popularfor numerical integrations.Equation (2.29) yields J2 = 2h(1 + A + ~A 2). Substituting this in (2.28) yields h N/2-1 h N/2-1 1= - L (61 + 6A + A 2)fzk = - L (flk+2 + 4flk+l 3 k=O 3 k=O h = "3 (fo + 4fl + 2fz + 4/3 + ... + 4fN-l + fN).

+ flk)

t

Simpson's three-eighths ruie for numerical integration

(2.31)

It is understood, of course, that N is an even integer.The factor gives this method its name. For Simpson's three-eighths rule, we set a = 3, retain terms up to A 3, and use Equation (2.29) to obtaio J3

= 3h(1 + ~A + ~A 2 + fA 3) = 3: (81 + l2A + 6A 2 + A 3).

76

2. OPERATOR ALGEBRA

Substituting in (2.28), we get 3h N/3-t 1=(81+12A+6A z+A 3) f3k 8 k=O 3h N/3-t = 8" ~ (f3k+3 + 3f3k+2 + 3f3k+t + 13k).

L

(2.32)

2.6.4. Example. Letus use Simpson's one-third rulewithfour intervals to evaluate the

familiar integrall

= fJ eX dx. With h =

0.25 and N

= 4, Equation (2.31) yields

1 '" 0.25 (I + 4e O.25 + 2e0.5 + 4e O.75 + e) = 1.71832. 3

This is very close to the"exact" result e - I '" 1.71828.

III

2.7 Problems 2.1. Considera linear operatorTon a finite-dimensional vector spaceV. Showthat there exists a polynomial P such that P (T) = O. Hint: Take a basis B = {Iai )}!:I aod consider the vectors {Tk lal) }~O for large enough M aod conclude that there exists a polynomial PI(n such that PI (T) lal) = O. Do the same for laz), etc. Now take the product of all such polynomials. 2.2. Use mathematicalinduction to show that [A, Am] = O.

2.3. For D aod T defined in Exarople 1.3.4: (a) Show that [D, T] = 1. (b) Calculate the linear traosformations D3T3 aod T3D3.

2.4. Consider three linear operators LI, Lz, aod L3 satisfying the commutationrelations [LI, Lz] = L3, [L3, Ltl = Lz, [Lz, L3] = LI, aod define the new operators L± = LI ± iLz. (a) Show that the operator LZ '" + ~ + L~ commutes with Lk, k = 1,2,3. (b) Show that the set {L+, L, L3} is closedundercommutation, i.e., the commutator of aoy two of them cao be written as a linear combinationof the set. Determine these commutators. (c) Write LZ in terms of L+, L_, aod L3.

Li

2.5. Prove the rest of Proposition 2.1.11. 2.6. Show that if [[A, B], A] = 0, then for every positive integer k, [Ak, B] = kAk-I[A, B].

Hint: First prove the relation for low values of k; then use mathematicalinduction.

2.7 PROBLEMS

2.7. Show that for 0 and T defined in Example 1.3.4, [Tk, OJ = _kTk - 1•

[ok, T]

=

kok-1

77

and

2.8. Evaluate the derivative of H- I (t) in terms of the derivative of H(t).

2.9. Show that for any ct, fJ

E

lR and any H E ,c,(V), we have

2.10. Show that (U + T)(U - T)

= U2 -

T2 if and only if [U, T]

= O.

2.11. Prove that if A and B are hermitian, then i [A, BJ is also hermitian.

2.12. Find the solution to the operator differential equation

dU

iii =

tHU(t).

Hint: Make the change of variable y = t 2 and use the result of Example 2.2.3.

2.13. Verify that

~H3 = dt

(dH) H2 dt

+ H (dH) H + H2 (dH). dt

dt

2.14. Show that if A and B commute, and !(A) and g(B) also commute.

!

and g are arbitrary functions, then

2.15. Assuming that US, T], T] = 0 = [[5, T], 5], show that [5, exp(tT)] = t[S, T] exp(tn. Hint: Expand the exponential and use Problem 2.6.

2.16. Prove that exp(HI

+ H2 + H3) =

exp(Hj ) exp(H2) exp(H3)

. exp{-!([HI, H2J

+ [HI, H3] + [H2, H3])}

provided that HI, H2, and H3 commute with all the commutators. What is the generalization to HI + H2 + ...+ Hn?

2.17. Denoting the derivative of A(t) by

A, show that

d . . -[A, B] = [A, BJ + [A, B].

dt

2.18. Prove Theorem 2.3.2. Hint: Use Equation (2.8) and Theorem 2.1.3.

2.19. Let A(t) es exp(tH)Ao exp( -tH), where H and An are constant operators. Show thatdA/dt = [H, A(t)]. What happens when H commutes with A(t)?

78

2. OPERATOR ALGEBRA

If}, Ig) E C(a, b) with the additional property that f(a) = g(a) =

2.20. Let

feb)

= g(b) = O. Show

that for such fuuctions, the derivative operator D is anti-hermitian, The inner product is defined as usual:

UI g) Heisenberg uncertainty principle

es

f

!*(t)g(t) dt.

2.21. In this problem, you will go through the steps of proving the rigorous statement of the Heisenberg nncertainty principle. Denote the expectation (average) value of an operator A in a state 11jI} by Aavg. Thus, Aavg = (A) = (1jI1 A 11jI). The uncertainty (deviation from the mean) in state 11jI) of the operator A is given by LlA = J«A - AavgP) = J(1jI1 (A - Aavg1P 11jI)·

(a) Show that for any two hermitian operators A and B, we have

Hint: Apply the Schwarz inequality to an appropriate pair of vectors. (b) Using the above and the triangle inequality for complex numbers, show that 1(1jI1 [A, B] 11jI)1 2 ::: 4 (1jI1 A 2 11j1) (1jI1 B 2 11j1) .

(c) Define the operators A' = A - "1, B' = B - fJ1, where" and fJ are real numbers. Show that A' and B' are hermitian and [A', B'] = [A, B]. (d) Now use all the results above to show the celebrated uncertainty relation (M)(LlB) ~

! I(1jI I [A, B] 11jI}1.

What does this reduce to for position operator x and momentum operator p if [x, p] = i1i? 2,22. Show that U = exp A is unitary if and only if A is anti-hermitian, 2.23. Find rt for each of the following linear operators. (a) T : JR2 -+ ]R2 given by

(b) T: ]R3 -+ ]R3 given by

T(~) = (;x~2:;~).

z -x+2y+3z

(c) T : ]R2 -+ ]R2 given by T

(X) = y

II)

(xcosll- y sin \xsmll+ycosll'

2.7 PROBLEMS

79

where 0 is a real number. What is TtT? (d) T : 1(;2 """* 1(;2 given by T

(a

1)

a'2

(e) T : 1(;3

=

(al -+ia2 ) • 1CXl

(X2

"""* 1(;3 given by

2.24. Show that if P is a (hermitian) projection operator, so are (a) 1 - P and (b) UtpU for any unitary operator U. 2.25. For the vector

(a) Find the associated projection matrix, Pa. (b) Verify that Pa does project an arbitrary vector in 1(;4 along la). (c) Verify directly that the matrix 1 - Pais also a projection operator.

=

=

2.26. Let lal) '" al (I, I, -I) and la2) sa 82 (-2, I, -I). (a) Construct (in the form of a matrix) the projection operators PI and P2 that project onto the directions of lal) and la2), respectively. Verify that they are indeed projection operators. (b) Construct (in the form of a matrix) the operator P = PI + P2 and verify directly that it is a projection operator. (c) Let P act on an arbitrary vector (x, y, z). What is the dot product of the resulting vector with the vector 8 I x 82? What can you say about P and your conclusion in (b)? 2.27. Let p(m) = L~=llei) (eil be a projection operator constructed out of the firstm orthonormal vectors of the basis B = {lei) 1f:,1of "17. Show that p(m) projects into the subspace spanned by the first m vectors in B. 2.28. What is the length of the projection of the vector (3, 4, -4) onto a line whose parametric equation is x = 21 + I, Y = -I + 3, z = 1 -I? Hint: Find a unit vector in the direction of the line and construct its projection operator. 2.29. The parametric equation of a line L in a coordinate system with origin 0 is

x=2/+1,

y = 1

+ I,

Z=-2/+2.

80

2. OPERATOR ALGEBRA

A point P has coordinates (3, -2, 1). (a) Using the projection operators, find the length of the projection of 0 P on the line L. (b) Find the vector whose beginning is P and ends on the line L and perpendicnlar to L. (c) From this vector calculate the distance from P to the line L.

2.30. Let the operator U : 1(;2 -> 1(;2 be given by U

(al) = (i Jz - i ~) a

~+.!!Z...

..(i..(i

2

Is U unitary? 2.31. Show that the product of two unitary operators is always unitary, but the product of two hermitian operators is hermitian if and only if they commute. 2.32. Let 5 be an operator that is both unitary and hermitian. Show that (a) 5 is involutive (i.e., 52 = 1). (b) 5 = p+ - P-, where p+ and P- are hermitian projection operators. 2,33. Show that when the forward difference operator is applied to a polynomial, the degree of the polynomial is reduced by I. (Hint: Consider x n first.) Then show that a n+1 annihilates all polynomials of degree n or less. 2.34. Show that 6 n+1 and V n+1 annihilate any polynomial of degree n. 2.35. Show that all of the finite-difference operators commute with one another. 2.36. Verify the identities

= 6E 1/ 2 = a, v + a = 2p,6 = E - E- 1, VE

E- 1/2

= P, _ (6/2).

2.37. By writing everything in terms of E, show that 6 2 = 2.38. Write expressions for E1/ 2 , 2.39. Show that

o = ~ sinh-I

a,

V, and p, in terms of 6.

m.

2.40. Show that

0 2 = :2 (a2 - a 3 + ga4 _ and derive Equation (2.27).

a - v

~a5 -

... )

=

a V.

2.7 PROBLEMS

81

2.41. Find an expression for Ja in powers of a. Retain all terms np to the foorth power. 2.42. Show that for a = 2, the third power of a disappears in Equation (2.29). 2.43. Evaluate the following integrals numerically, using six subintervals with the trapezoidal rule, Simpson's one-third rule, and Simpson's three-eighths rule. Compare with the exact result when possible.

[o ,

e- X dx.

(c)

(d) [ 1 -dx. 1 x

(e) 14inXdx.

(f)

(g) [

(h)

[,

(i)

1 5

(a)

x

_1

3dx.

dx 1 + x 2'

(b)

0

xe" dx.

f2 0

xei cos x dx.

{'e' 1

1

sin x dx.

1

eX tan x dx.

Additional Reading 1. Axler, S. Linear Algebra Done Right, Springer-Verlag, 1996. 2. Greub, W. Linear Algebra, 4th ed., Springer-Verlag, 1975. 3. Hildebrand, F. Introduction to Numerical Analysis, 2nd ed., Dover, 1987. Uses operator techniques in numerical analysis. It has a detailed discussion of error analysis, a topic completely ignored in our text.

3

_

Matrices: Operator Representations

So far, our theoretical investigation has been dealing mostly with abstract vectors and abstract operators. As we have seen in examples and problems, concrete representations of vectors and operators are necessary in most applications. Such representations are obtained by choosing a basis and expressing all operations in terms of components of vectors and matrix representations of operators.

3.1 Matrices Letus choose a basis Bv = {Iai) 1~1 ofa vector space VN, and express an arbitrary vector Ix) in this basis: [r) = L~I ~i lai). We write

(3.1)

representation of vectors

and say that the column vector x represents Ix) in Bv . We can also have a linear transformation A E L (VN, W M) act on the basis vectors in Bv to give vectors in the M-dimensional vector space WM: IWk) = A lak). The tatter can be written as a linear combination of basis vectors Bw = {Ibj) If=j in W M : M

IWI) = L"jllbj), j=l

M

IW2) = L"j2Ibj), j=l

M

IWN) = L"jN Ibj)' j=1

Note that the components have an extra subscript to denote which of the N vectors {I Wi ) 1~1 they are representing. The components can be arranged in a column as

83

3.1 MATRICES

before to give a representation of lbe corresponding vectors:

The operator itself is determined by the collection of all these vectors, i.e., by a matrix. We write this as "' 11 "'21

A= (

representation of operators

.;

(3.2)

a~2

and call Albe matrix representing A in bases Bv and Bw. This statement is also sunnnarized symbolically as M

Alai) = I>jilbj) ,

i = 1,2, ... ,N.

(3.3)

j~1

We lbus have the following mle:

3.1.1. Box. To find the matrix A representing A in bases Bv = (Iai) 1~1 and Bw = lib j) 1f=1' express Alai) as a linear combination of the vectors in Bw. The components form the i th column of A.

Now consider lbe vector Iy) = A Ix) in WM. This vector can be written in two ways: On the one hand, Iy) = Ef=1 ryj Ibj ) . On the other hand,

IY) =Alx) = A

N

N

;=1

;=1

I) lai) = L~iAlai)

Since Iy) has a unique set of components in lbe basis Bw, we conclude lbat N

ryj =

L"'ji~i' i=l

j = 1,2, ... ,M.

(3.4)

84

3. MATRICES: OPERATOR REPRESENTATIONS

This is written as

m ) (al1. ·

~2

( ··

~M

The operator TA associated with a matrix A

a21

a12 a22

(3.5)

..

aMI

aM2

in which the matrix multiplication rule is understood. This matrix equation is the representation of the operator equation Iy) = A [x) in the bases Bv and Bw. The construction above indicates that-once the bases are fixed in the two vector spaces-to every operator there corresponds a unique matrix. This uniqueness is the result of the uniqueness of the components of vectors in a basis. On the other hand, given an M x N matrix A with elements aij, one can construct a unique linear2Perator TA defined by its action on the basis vectors (see Box 1.3.3): TA lai) == Ljd aji Ibj ). Thus, there is a one-to-one correspondence between operators and matrices. This correspondence is in fact a linear isomorphism: 3.1.2. Proposition. The two vector spaces £.,(VN, W M) and JV(M x N are isomorphic. An explicit isomorphism is established only when a basis is chosen for each vector space, inwhichcase, anoperator is identified with its matrixrepresentation. Given the linear transformations A : VN -+ WM and B : WM -+ UK, we can form the composite linear transformation BoA : VN -+ UK. We can also choose bases Bv = {Iai)}~" Bw = {Ibi)}~" Bu = {Ici)};;', for V, W, and U, respectively. Then A, B, and BoA will be represented by an M x N, a K x M, and a K x N matrix, respectively, the latter being the matrix product of the other two matrices. Matrices are determined entirely by their elements. For this reason a matrix A whose elements are (.lll, 0!12, ... is sometimes denoted by (aij). Sintilarly, the elements of this matrix are denoted by (A)ij. So, on the one hand, we have (aij) = A, and on the other hand (A)ij = ai], In the context of this

notation, therefore, we can write (A + B)ij

= (A)ij + (B)ij

=}

(aij

+ fJij) = (aij) + (fJij),

(yA)ij = y(A)ij

=}

y(aij) = (yaij),

(O)ij = 0, (1)ij = 8ij.

A matrix as a representation of a linear operator is well-defined only in reference to a specific basis. A collection ofrows and columns ofnumbers by themselves have no operational meaning. When we manipulate matrices and attach meaning to them, we make an unannounced assumption regarding the basis: We have the standard basis of en (or JRn) in mind. The following example should clarify this subtlety.

3.1 MATRICES

85

3.1.3. Example. Let us fiud the matrix represeutatiou of the liuear operator A E £, (lR. 3), given by

X) (X-Y+2Z) (z = 3x-z 2y+z

A Y

(3.6)

in thebasis

There is a tendency to associate thematrix

-1

o 2

withtheoperator A. Thefollowingdiscussionwill showthatthisis false.Toobtain thefirst columnof the matrix representing A, we consider

So, by Box 3.1.1, the first column of the matrix is

The othertwo columns are obtained from

giving the secondandthethird columns, respectively. The whole matrix is then

2-5~)t . 1

o

2

As long as all vectors are represented by columnswhose entries are expansion coefficients of thevectors in B, A andAare indistinguishable. However, the actionof A on the column

86

3. MATRICES: OPERATOR REPRESENTATIONS x

vector (Y ) willnot yieldtheRHS ofEquation(3.6)!Althoughthis is notnsuallyemphasized, z the columnvectoron theLHS of Equation (3.6) is reallythe vector

whichis an expansion in terms of the standard basis of]R3 rather than in tenus of B. x

Wecan expand A ( Y) in terms of B. yielding

z

A

X) (X - Y + 2Z) (Yz = 3x-z 2y+z = (2x -

b)

(i)

+(-x+ h+ 2z)

G) +

(x+

b

-z)

(D.

Thissays that in thebasis B thisvector has therepresentation

(3.7)

Similarly,

X

(Y ) is represented by z

W)) z

= ( B

t~ ~~: ~; ).

(3.8)

-~x + ~y + ~z

Applying A to the RHS of (3.8) yieldsthe RHS of (3.7),as it should.

rank ofa matrix

Given any M X N matrix A, an operator TA E L (VN, W M) can be associated with A, and one can construct the kernel and the range of TA. The rank of TA is called the rank of A. Sioce the rank of an operator is basis independent, this definition makes sense. Now suppose that we choose a basis for the kernel ofTA and extend it to a basis of V. Let VI denote the span of the remaining basis vectors. Similarly, we choose a basis for TA (V) and extend it to a basis for W. In these two bases, the M x N matrix representiog TA will have all zeros except for an r x r submatrix, where r is the rank of TA. The reader may verify that this submatrix has a nonzero detenninant. In fact, the submatrix represents the isomorphism betweeu VI and TA (V), and, by its very construction, is the largest such matrix. Since the determinant of an operator is basis- iudependent, we have the followiog proposition:

3.1.4. Proposition. The rank ofa matrix is the dimension ofthe largest (square) submatrix whose determinant is not zero.

3.2 OPERATIONS ON MATRICES

87

3.2 Operations on Matrices transpose ofa matrix

There are two basic operations that one can perform on a matrix to obtain a new one; these are transposition and complex conjugation. The transpose of an M x N matrix A is an N x M matrix At obtained by interchanging the rows and columns of A: (At)ij = (A)ji,

(3.9)

or

The following theorem, whose proof follows immediately from the definition oftranspose, summarizes the important properties ofthe operation of transposition.

3.2.1. Theorem. Let A and B be two (square) matrices. Then (a) (A + B)t = At

symmetric and antisymmetric matrices

orthogonal matrix complex conjugation

(b) (AB)t = Bt At,

+ Bt,

Of special interest is a matrix that is identical to its transpose. Such matrices occur frequently in physics and are called symmetric matrices. Similarly, antisymmetric matrices are those satisfying At = -A. Any matrix A can be written as A = ~(A + At) + ~(A - At), where the first term is symmetric and the second is anti symmetric. The elements of a symmetric matrix A satisfy the relation a ji = (At)ij = (A)ij = aij; i.e., the matrix is symmetric under reflection through the main diagonal. On the other hand, for an antisymmetric matrix we have a ji = -aij. In particular, the diagonal elements of an antisymmetric matrix are all zero. A (real) matrix satisfying At A = AAt = 1 is called orthogonal. Complex conjugation is an operation under which all elements of a matrix are complex conjugated. Denoting the complex conjugate of A by A*, we have (A·)ij = (A)ij, or (aij)* == (a~). A matrix is real if and only if A* = A. Clearly,

Wr=A

hermitian conjugate

(c) (At)t = A.

.

Under the combined operation of complex coujugation and transpositiou, the rows and columns of a matrix are interchanged and all of its elemeuts are complex conjugated. This combined operatiou is called the adjoint operation, or hermitian conjugation, and is denoted by t, as with operators. Thus, we have At

(At)ij

= (At)* = (A*)t,

= (A)ji

or

(aij)t

= (aji)'

(3.10)

Two types of matrices are important enough to warrant a separate definition. hermitian and unitary matrices

3.2.2. Definition. A hermitian matrix H satisfies Ht = H, or, in terms ofelements, ~ij

= tl ji- A unitary matrix U satisfies UtU = uut = 1, or, in terms of elements,

"N L....k=l

"L...k=l N* lLiktL* jk = J.Lkif.Lkj

, = Vij·

Remarks: It follows immediately from this definition that I. The diagonal elements of a hermitian matrix are real.

88

3. MATRICES: OPERATOR REPRESENTATIONS

2. The kth column of a hennitiau matrix is the complex coujugate of its kth row, and vice versa. 3. A real hermitiau matrix is symmetric. 4. The rows of au N x N unitary matrix, when considered as vectors in eN, form

diagonal matrices

au orthonormal set, as do the columns. 5. A real unitary matrix is orthogonal. It is sometimes possible (aud desirable) to trausfonn a matrix into a form in which all of its off-diagonal elements are zero. Such a matrix is called a diagonal matrix. A diagonal matrix whose diagonal elements are {Akl~1 is denoted by diag(AI, A2, ... , AN)'

3.2.3. Example. In this example,we derive a useful identityfor functions of a diagonal matrix.Let D = diag(AI,1.2, ... , An) be a diagonalmatrix, and [(x) a functionthat bas a Taylor series expansion f (x) = L~o akxk. The samefunction of D canbe written as

In words, thefunction ofadiagonal matrix is equal toadiagonal matrix whoseentries are the samefunction of the corresponding entries of the original matrix. ill the abovederivation, we used the followingobviousproperties of diagonal matrices:

adiag(AJ,A2, ... ,An) = diag(aAI,aA2, ... ,aAn), diag(AI,1.2, , An) + diag("'I, "'2, , "'n) = diag(AI +"'1, ... , An + Wn), diag(AI,1.2, , An) . diag("'I. "'2 , "'n) = diag(AI"'I, ... , An"'n).

III

3.2.4. Example. (a) A prototypical symmetric matrix is that of the moment of inertia encountered in mechanics. The ijth eleIIlent of this matrix is defined as Iij es fff P(XI. X2, X3)XiXj dV, where Xi is the ilb Cartesiancoordinateof a point in the distribution of mass described by the volume density P(Xlo xz, X3). It is clear that Iij = Iji. or I = It. The moment of inertia matrix canbe represented as

It has six independent elements. (b) An exampleof an anti symmetric matrix is the electromagnetic field tensor given by

F= (~3

-82 -EI

Pauli spin matrices

-83

o

81 -E2

82 -81 0 -E3

(c) Examples of hermitian matrices arethe2 x 2 Pauli spin matrices:

3.3 ORTHONORMAL BASES

0

cz = ( i Euler angles

89

-i) 0 .

(d) The most frequently encountered orthogonal matrices are rotations. One suchmatrix represents therotation of a 3-dimensional rigidbodyin terms of Euler angles andis usedin mechanics. Attaching acoordinate system tothebody, ageneral rotation canbedecomposed intoa rotation of anglee about the z-axis, followed by a rotation of angle e about the new x-axis, followedbyarotation of angle1ft about thenew z-axis. Wesimply exhibit thismatrix in terms of theseangles andleave it to thereader to showthat it is indeedorthogonal. cos tp - sin 1/1 cos 8 sin cp sin 1/1 cos rp + cos 'if.r cos () sin rp

COS 1fr (

sin fJ sin rp

- cos 1fr sin q; - sin 1/1 cos 8 cos rp - sin "" sin cp + cos 1/1 cos ()cos qJ sine cos rp

sin1f.rsin9 ) - cos 1ft sinB . cosB III

3.3 Orthonormal Bases Matrix representation is facilitated by choosing an orthonormal basis B = (Ie;)}!:!. The matrix elements of an operator A can be found in such a basis by "multiplying" both sides of A lei) = I:f=t aki lek) on the left by {ej I:

or (3.11) We can also show that in an orthonormal basis, the ith component ;i of a vector is found by multiplying it by (ei I. This expression for;i allows us to write the expansion of [x) as N =}

1 = I>j}(ejl,

(3.12)

j~l

which is the same as in Proposition 2.5.3. Let us now investigate the representation of the special operators discussed in Chapter 2 and find the connection between those operators and the matrices encountered in the last section. We begin by calculating the matrix representing the hermitian conjugate of an operator T. In an orthonormal basis, the elements of this matrix are given by Equation (3.11), 7:;j = (eil T lej). Taking the complex conjugate of this equation, and using the definition of T" given in Equation (2.8), we obtain 7:ij = (eil T lej)* = (ej Irt lei), or (Tt)ij = 7:j;. This is precisely how the adjoint of a matrix was defined.

90

3. MATRICES: OPERATOR REPRESENTATIONS

Note how crucially this conclusion depends on the orthonormality of the basis vectors. If the basis were not orthonormal, we could not use Equation (3.11) on which the conclusion is based. Therefore,

3.3.1. Box. Only in an orthonormal basis is the adjoint ofan operator represented by the adjoint of the matrix representing that operator.

In particular, a hermitian operator is represented by a hermitian matrix only if an orthonormal basis is used. The following example illustrates this point. 3.3.2. Example. Consider the matrix representation of the hermitian operator H in a general-not orthononnal-basis B = {Iai) If::l . Theelementsofthe matrix corresponding to H aregiven by N

H lak) =

L

N

tl jk laj).

or

H la;) =

j=l

L tl ji laj}'

(3.13)

j=l

Taking the product of the first equation with
(ail H lak)* = (akl Ht lai) = (akl H lai)' So we have (akl H lai) = Ef=l ~jk (ajlai). On the other hand, muitiplying the second equation in (3.13) by (akl gives (akl H jai) = "L.7=1 Ylji {akl aj}' Theonly conclusion we candraw from thisdiscussion is

Ef~l ~jk (ajl ail = Ef~l ~ji (akl aj). Because litis equation does not say anything about each individual 1Jij. we cannot conclude, in general, that = 'YJji' However, if the lai)'s are orthonormal, then (ajl ail = 8ji and (akl aj) = 8kj. and we obtain

1]0

"L.f=l11jk(Jji

= "Lf=11Jji8icj, or rJik =

nu. as expected of a hermitian matrix.

II

Similarly, we expect the matrices representing unitary operators to be unitary only if the basis is orthonormal. This is an inunediate consequence of Equation (3.10), hut we shall prove it in order to provide yet another example of how the completeness relation, Equation (3.12), is used. Since UU t = 1, we have (eil UU t lej) = (eil1Iej) = 8ij. We insert the completeness relation 1 = Lf=l lek) (ekl between U and ut on the LHS: (eil U

(~ lekHekl) Ut lej) = ~ (eil ~ lek) (ekl ~t lej) = 8ij. =f.£ik

=f.£jk

This equation gives the first half of the requirement for a unitary matrix given in Definition 3.2.2. By redoing the calcnlation for UtU, we conld obtain the second half of that requirement.

3.4 CHANGE OF BASIS AND SIMILARITY TRANSFORMATION

91

3.4 Change of Basis and Similarity Transformation !tis often advantageous to describe a physical problem iu a particular basis because it takes a simpler form there, but the general form of the result may still be of importance. In such cases the problem is solved in one basis, and the result is transformed to other bases. Let us investigate this point in some detail. Given a basis B = (Iai)}!:!' we can wtite an arbitrary vector la) with components {al,a2, ... ,aN} in B as la) = L!:lai lai). Now suppose that we change the basis to B' = (Iai)}f=l' How are the components of la) in B' related to those in B? To answer this question, we wtite lai) in terms of B' vectors, lai) = Lf=l Pji lai), and substitute for lai) in this expansion of la), obtaining la) = L!:l a; Lf=l Pji lai) = Li,j aiPji lai)· If we denote the jth component of la) in B' by ai, then this equation tells us that N

otj =

LP

jiotj

for j = 1,2, "" N.

(3.14)

i=l

If we use a', R, and a, respectively, to designate a column vector with elements ai, an N x N matrix withelements Pit- and a columnvectorwith elements ai, then Equation (3.14) can be wtitten in matrix form as

(a~

~~) cp~: ~~

basistransformation matrix

-

~l P~2

••

a~'

or

a' = Ra.

(3.15)

The matrix R is called the basis transformation matrix. !t is invertible because it is a linear transformation that maps one basis onto another (see Theorem 2.1.6). What happens to a matrix when we transform the basis? Consider the equation Ib) = A la), where la) and Ib) have components {ail!:l and (,Bil!:!' respectively, in B. This equation has a corresponding matrix equation b = Aa. Now, if we cbaoge the basis, the components of la) and Ib) will change to those of a' and b', respectively. We seek a matrix A' such that b' = A'a'. This matrix will clearly be the transform of A. Using Equation (3.15), we wtite Rb = A'Ra, or b = R-1A'Ra. Comparing this with b = Aa and applying the fact that both equations hold for arbitrary a and b, we conclude that or

similarity transformation

~~~) (~~)

(3.16)

This is called a similarity transformation on A, and A' is said to be similar to A. The transformation matrix R can easily be found for orthonormal bases B = (Jei)}!:l and B' = (Ie;)}!:!. We have lei) = Lf=l Pki le~). Multiplying this

92

3. MATRICES: OPERATOR REPRESENTATIONS

equation by {ei I, we obtain N

(eilei)

N

= LPki (eileiJ = LPki 8jk
(3.17)

k~1

That is, 3.4.1. Box. To find the ijth element of the matrix that changes the components of a vector in the orthonormal basis B to those of the same vector in the orthonormal basis B', take the j th ket in B and multiply it by the ith bra in R'. To find the ijth element of the matrix that changes B' into B, we take the jth ket in B' and multiply it by the ith bra in B: pij = (ed ei). However, the matrix R' must be R- I, as can be seen from Equation (3.15). On the other hand, (pi)' (edei)' (eilei) Pji,Or

=

(R-I)lj

=

=Pji,

=

or

(R-I)jj

=Pji =(Rt)ij.

(3.18)

This shows that R is a unitary matrix and yields an important result. 3.4.2. Theorem. The matrix that transforms one orthonormal basis into another is necessarily unitary. From Equations (3.17) and (3.18) we have (Rt)ij = {ed ei}. Thus,

3.4.3. Box. To obtain the jth column of Rt, we take the jth vector in the new basis and successively "multiply" it by {ed for i = 1,2, ... , N.

In particular, if the original basis is the standard basis of

eN and

lei} is rep-

resented by a column vector in that basis, then the jth column of Rt is simply the vector lei}. 3.4.4. Example. In this example, we show that the similarity transform of a function of a matrix is the same function of the similarity transform of the matrix: R!(A)R- 1 = f (RAR-1). Theproofinvolvesinserting 1 = R- I Rbetweenfactors of Ain theTaylor series

expansion of !(A): R!(A)R- I = R

If. akAk) \.t=0

R- I = f;akRAkR-I =

k~O

f;akR~~-;R-I

k~O

k times 00

.

00

=I>k RAR-IRAR- I ... RAR- I = L ak (RAR-I)k = !(RAR-I). k~O

k=O

III

3.5 THE DETERMINANT

93

3.5 The Determinant An important concept associated with linear operators is the determinant, Determinants are usually defined in terms of matrix representations of operators in a particular basis. This may give the impression that determinants are basis dependent. However, we shall show that the value of the determinant of an operator is the same in all bases. In fact, it is possible to define determinants of operators without resort to a specific representation of the operator in terms of matrices (see Section 25.3.1).

Letus first introduce a permutation symbol eiliz ...iN' whichwill be usedextensively in this chapter. It is defined by (3.19) In other words, S'j'2 ...iN is completely antisymmetric (or skew-symmetric) under interchange of any pair of its indices. We will use this permutation symbol to define determinants. An immediate consequence of the definition above is that S'j'2 ...iN will be zero if any two of its indices are equal. Also note that f:iliz .. .iN is + 1 if (iI, ia. ... , iN) is an even permutation I (shuflling) of (1, 2, . '" N), and -I if it is an odd permutation of (1,2, ... , N).

3.5.1 Determinant of a Matrix determinant defined

3.5.1. Definition. The determinant is a mapping, det : MNxN --> C, given in term of the elements OIu ofa matrix A by N

detA

=

L

8iliz ...iNCXlit ... 00NiN'

il, ...,iN=l

Definition 3.5.1 gives detA in terms of an expansion in rows, so the first entry is from the first row, the second from the second row, and so on. It is also possible to expand in terms of columns, as the following theorem shows. 3.5.2. Theorem. The determinant ofa matrix A can be written as N

detA =

L

eiriz...iNCXitl ... otiNN·

il, ...,iN

Therefore, det A = det At. 1An even permutation means an even number of exchanges of pairs of elements. Thus, (2,3, 1) is an even permutation of (1, 2, 3), while (2, 1, 3) is an odd permutation. It can be shown (see Chapter 23) that the parity (evenness or oddness) of a permutation is well-defined; i.e., that although there may be manyroutes of reaching a permutation from a given (fiducial)

permutation via exchanges of pairs of elements, all such routes require eithereven numbers of exchanges or odd numbers of exchanges.

94

3. MATRICES: OPERATOR REPRESENTATIONS

Proof We shall go into some detail in the proof of only this theorem on determi-

nants to illustrate the manipulations involved in working with the B symbol. We shall not reproduce such details for the other theorems. In the equation of Definition 3.5.1, ii, i2, ... , iN are all different and form a permutation of (I, 2, ... ,N). So one of the a's must have I as its second index. We assume that it is the h th term; that is, iiI = I, and

We move this term all the way to the left to get aiJIali! . .. aNiN' Now we look for the entry with 2 as the second index and assume that it occurs at the hth position; that is, i jz = 2. We move this to the left, next to a iI 1, and wtite (ljt1(ljz2(llil aNiN0 We continue in thisfashion untilwe get (lhl(ljz2· .. oriNN. Since h, iN is really a reshuffling of it, i2, ... , iN, the surnmationindices can be changed to h, h. ... , l«, and we can wtite2

n. , N

detA =

"I:

Bili2 ...iN(1hl ... ajNN.

il,···,jN=l

If we can show that Biliz ...iN = 8 hjz ...in» we are done. In the equation of the theorem, the sequence of integers (ii, i2, ... , is) is obtained by some shuffling of (I, 2, ... , N). What we have done just now is to reshuffle (iI, tz. ... , iN) in reverse order to get back to (1, 2, ... , N). Thus, if the shuflling in the equation of the theorem is even (odd), the reshuffling will also be even (odd). Thus, Biti, ...iN = Biih...is» and we obtain the first part of the theorem: N

detA =

L

8hjz ...jN(lhl ... ajNN.

h,···,iN=l

For the second part, we simply note that the rows of At are coluums of A and vice versa. D

3.5.3. Theorem. Interchanging two rows (or two columns) of a matrix changes the sign of its determinant. Proof The proof is a simple exercise in permutation left for the reader.

D

An immediate consequence of this theorem is the following corollary.

3.5.4. Corollary. The determinant ofa matrix with two equal rows (or two equal columns) is zero. Therefore, one can add a multiple ofa row (column) to another row (column) ofa matrix without changing its determinant. 2The E symbolin the sumis not independent of the j's. although it appears without suchindices. In reality, the i indicesare "functions" of the j indices.

3.5 THE DETERMINANT

95

Since every term of the determinant in Definition 3.5.1 contains one and only

one elementfrom each row, we can write N

detA = CXilAil +cxi2Ai2

+ ... +otiNAiN = L:CiijAij, j~l

cofactor ofan element of a matrix

where Aij contaios products of elements of the matrix A other than the element aii' Since eachelementof aroworcolumnoccursatmost once in each term of the expansion, Aij cannot contain any element from the ithrow or the jth column, The quantity Aij is called the cofactor of aij, and the above expression is known as the (Laplace) expansion ofdetA by its ith row. Clearly, there is a similar expansion by the ith column of the determinant, which is obtained by a similar argument usiog the equation of Theorem 3.5.2. We collect both results io the followiog equation. N detA = L:cxijAij )=1

=

N L:ajiAjio

(3.20)

)=1

Vandermonde, Alexandre-Thieophile, alsoknownas Alexis, Abnit, andCharles-Auguste Vandermonde (1735-1796) had a father, a physician who directed his sickly son toward a musicalcareer. An acquaintanceship with Fontaine, however, so stimulated Vandennonde that in 1771 he was elected to the Academic des Sciences, to which he presented four mathematical papers (histotal mathematical production) in 1771 ~1772. Later, Vandennond.e wrote several papers on harmony, and it was said at thattime that musicians considered Vandermonde to be a mathematician andthat mathematicians viewed him as a musician. Vandermonde's membership in the Academy led to a paper on experiments withcold, made with Bezout and Lavoisier in 1776, and a paper on the manufacture of steel with Berthollel and Monge in 1786. Vandermonde became an ardent and active revolutionary, being such a close friend of Monge thathe was termed "femme de Monge." He was a member of the Commune of Paris andthe club of the Jacobins. In 1782 he was director of the Conservatoire des Arts et Metiers and in 1792, chief of the Bureau de 1'Habillement des Armies. Hejoinedin thedesignof a coursein politicaleconomyfortheEcole Normale andin 1795 was named a member of the Institut National. Vandermonde is best known for the theory of determinants. Lebesgue believed that the attribution of determinant to Vandermonde was due to a misreading of his notation. Nevertheless, Vandermonde's fourth paper was the first to give a connected exposition of determinants, because he (1) defined a contemporary symbolismthat was more complete, simple, and appropriate than that of Leibniz; (2) defined determinants as functions apart from thesolution of linear equations presented by Cramer butalso treated by Vandermonde; and(3) gave a number of properties of thesefunctions, such as thenumber andsigns of the terms andtheeffect of interchanging two consecutive indices(rowsor columns), whichhe usedto showthata determinant is zero if two rowsor columnsareidentical. Vandermonde's real andunrecognized claim to fame was lodged in his first paper, in which he approachedthe generalprohlem of the solvabilityof algebraiceqoationsthrough a study of functions invariant under permutations of the roots of the equations. Cauchy

96

3. MATRICES: OPERATOR REPRESENTATIONS

assigned priority in this to Lagrange and Vandermonde. Vandermonde read his paper in November 1770, but he did not become a member of the Academy until 1771, and the

paper was not publisheduntil 1774.AlthoughVandennonde's methodswere close to those later developed by Abel and Galois for testing the solvability of equations, and although his treatment of the binomial equation x n - 1 = 0 could easily have led to the anticipation of Gauss's results on constructible polygons, Vandennonde himself did not rigorously or

completelyestablishhis results, nor did he see the implicationsfor geometry. Nevertheless, Kroneckerdates the modern movement in algebra to Vandermonde's 1770 paper. Unfortunately, Vandennonde's spurt of enthusiasm and creativity, which in two years produced four insightful mathematical papers at least two of which were of substantial importance, was quickly diverted by the exciting politics of the time and perhaps by poor

health.

3.5.5. Proposition. Ifi

i' k, then L.f~1 aijAkj = 0 = L.f~1 ajiAjk.

Proof Consider the matrix B ohtained from A hy replacing row k hy row i (row i remains unchanged, of course). The matrix B has two equal rows, and its determinant is therefore zero. Now, if we ex~and det B by its kth row according to Equation (3.20), we obtain 0 = det B = L.j=1 t!kjBkj. But the elements of the kth row of B are the elements of the ith row of A; that is, {Jkj = ai], and the cofactors of the kth row of B are the same as those of A, that is, Bkj = Akj. Thus, the first equation of the proposition is established. The second equation can be established using expansion by columns. D

minor ofa matrix

A minor of order N - I of an N x N matrix A is the determinant of a matrix obtained by striking out one row and one column of A. If we strike out the ith row and jth column of A, then the minor is denoted by Mij. 3.5.6. Theorem. Aij = (-l)'+j Mij. Proof The proofinvolves separating all from the rest of the terms in the expansion of the determinant, The unique coefficient of all is All by Equation (3.20). We can show that it is also MIl by examining the e expansion of the determinant and performing the first sum. This will establish the equality A II = MIl. The general equality is obtained by performing enough interchanges of rows and columns of the matrix to bring aij into the first-row first-column position, each exchange introducing a negative sign, thus the (-l)i+j factor. The details are left as an exercise. D

The combination of Equation (3.20) and Theorem 3.5.6 gives the familiar routine of evaluating the determinant of a matrix.

3.5.2 Determinants of Products of Matrices One extremely useful property of determinants is expressed in the following theorem.

3.5 THE DETERMINANT

97

3.5.7. Theorem. det(AB) = (detA)(detB)

Proof The proof consists in keeping track of index shuffling while rearranging the order of products of matrix elements. We shall leave the details as an exercise.

D

3.5.8. Example. Let 0 and U denote, respectively, an orthogonal and a unitary n x n matrix; that is, 00' = 0'0 = 1, and UU t = utu = 1. Taking the detenoinant of the first equation and nsing Theorems 3.5.2 and 3.5.7. we nbtain (det O)(det a') = (det 0)2 = det 1 = 1. Therefore, for an orthogonal matrix, we get detO = ±1. Orthogonal transformations preserve a real inner product. Among such transformations are the so-called inversions, which, in their simplest form, multiply a vector by -1. In three dimensions this corresponds to a reflection through the origin. The matrix associated with this operation is -1:

o -1

o which has a determinant of -1. This is a prototype of other, more complicated, orthogonal transformations whose determinants are -1. The other orthogonal transformations, whose determinants are + I, are of special interest because they correspond to rotations in three dimensions. The set of orthogonal transformations in n dimensions having the determinant +1 is denoted by SO(n). These transformations are special because they have the mathematical structure of a (continuous) group, which finds application in many areas of advanced physics. We shall come back to the topic of group theory later in the book. We can obtain a similar result for unitary transformations. We take the determinant of both sides of UtU = 1:

det(U*)' det U = det U* det U = (det U)*(det U) = Idet UI 2 = I. Thus, we can generally write det U = e ia, with a E JR.. The set of those transformations with a = Ofonos agronp to which 1 belongs and that is denoted by SU(n). This gronphas found applications in the attempts at unifying the fundamental interactions. II

3.5.3 Inverse of a Matrix One of the most useful properties of the determinant is the simple criterion it gives for a matrix to be invertible. We are ready to investigate this criterion now. We combine Equation (3.20) and the content of Proposition 3.5.5 into a single equation,

N N LClijAkj = (detA)8ik = LCljiAjk, )=1

j=l

(3.21)

98

3. MATRICES: OPERATOR REPRESENTATIONS

andconstruct a matrix CA, whose elementsarethe cofactors of the elements of the matrix A:

(3.22)

Then Equation (3.21) can be written as N

N

j=l

j=l

I>ij (C~)jk = (detA)8ik = L (C~)kj «s. Of,

in matrix form, as AC~

inverse ofamatrix

= (det A)1 = C~A.

(3.23)

3.5.9. Theorem. The inverse ofa matrix (if it exists) is unique. The matrix A has an inverse if and only ifdetA i' O. Furthermore,

C'

A-t = _A_ det A'

(3.24)

where C A is the matrix ofthe cofactors of A. Proof. Let B and C be inverses of A. Then B = (CA) B = C (AB) = C. '-.r-'

'-.r-'

=1

=1

For the second part, we note that if A has an inverse B, then AB = 1

=}

detAdetB

= det 1 = I,

whence det A i' O. Conversely, if det A i' 0, then dividing both sides of Equation D (3.23) by detA, we obtain the unique inverse (3.24) of A. The inverse of a 2 x 2 matrix is easily found:

a (e

b)-t d

d -b)

I ( =ad-be -e

a

(3.25)

ifad - be i' O. There is a more practical way of calculating the inverse ofmatrices. In the following discussion of this method, we shall confine ourselves simply to stating a couple of definitions and the main theorem, with no attempt at providing any proofs. The practical utility of the method will be illusttated by a detailed analysis of examples.

3.5 THE DETERMINANT

elementary row operation

99

3.5.10. Definition. An elementary row operation on a matrix is one of the following: (a) interchange oftwo rows ofthe matrix, (b) multiplication ofa row by a nonzero number, and (c) addition ofa multiple ofone row to another. Elementary column operations are defined analogously.

triangular, or row-echelon form of a matrix

3.5.11. Definition. A matrix is in triangular, or row-echelon.form the following three conditions:

if it satisfies

1. Any row consisting ofonly zeros is below any row that contains at least one nonzero element. 2. Goingfrom left to right, the first nonzero entry ofany row is to the left ofthe first nonzero entry ofany lower row. 3. The first nonzero entry of each row is 1. 3.5.12. Theorem. For any invertible n x n matrix A, • The n x 2n matrix (All) can be transformed into the n x 2n matrix (IIA-I) by means ofa finite number of elementary row operations. 3

• if (All)

is transformed into (liB) by means of elementary row operations, then B = A-I

A systematic way of transforming (All) into (I lA-I ) is first to bring A into triangular form and then eliminate all nonzero elements of each column by elementary row operations. 3.5.13. Example. Let us evaluate the inverse of

A=

G: =D·

G -I II 0 0)

We start with 2

o

I I

-2 -I

0 0

I 0

0 I

== M

and apply elementary row operations to Mto bring the left half of it into triangular form. If we dennte the klb row by (k) and the three operatinns of Definition 3.5.10, respectively, by (k) (j), ark), and ark) + (j), we get

'*

M

(I -I II (I -I II ,0

-2(1)+(3)

-------+

-!(3)

0 0

2

I

-2

0

0

-3

I

-2

2 I 0

-2 I

0 2/5

0

I

0

0 I -3/5

0)o I

I

, o

2 I

3(2)+(3) ( o 0

-I II -2

0

oI

-5

-2

3

0) 0 I

o o ) =M' -1/5

3The matrix (AI1) denotes the n x 2n matrix obtained by juxtaposing the n x n unit matrix to the right of A. It can easily be shown that if A, 8, and C are n x n matrices, then A(BIC) = (ABIAC).

100

3. MATRICES, OPERATOR REPRESENTATIONS

The left half of M' is in triangular form. However, we want all entries above any 1 in a columnto be zero as well, .i.e., we want the left-hand matrix to be 1. We can do this by appropriate use of type 3 elementary row operations: M'

0

0 I 0

3 II0 -2 I 2/5

G

0 I 0

-2 I

c 0

) -2(2)+(1)

, -3(3)+(1)

2(3)+(2)

G

o

0 I 0

1-1/5 0 2/5

I

-1/5 I -3/5

-1/5 -1/5 -3/5

o ,-1/5 o 4/5 2/5

-tJ

-2 I -3/5

3/5 )

-~/5

3/5 ) -2/5 -1/5

The right half of the resultingmatrix is A-I.

11II

3.5.14. Example. It is instructive to start with a matrix thatis not invertible andshow that it is impossibleto turn it into 1 by elementary rowoperations. Consider the matrix

B=

(i

-I -2 5

-I



Let us systematically bring it into triangular fonn:

M=

(i

-I -2 5

-I

, -2(1)+(2)

'

-(2)+(3)

1 0 31 I 0 I o 0 0 -2 3 5

(~I

C 00

-2 3 0

0)o I

II 10I o 0

II I 0I o -I

C

,2 (1)<>(2) -I

I -2 0 I -2 3

0)o I

-2 -I 5

-------+ (1)+(3)

C

o 0) C I

--+ ~(2)

0 0

3I 10I o 0

0 0

-2 I 0

-2 3 3

I 0 0

D

II 10I I 0

I I 1/3 0 1/3 o -I

I -2 I

D

I -2/3 3

D

The matrix B is now in triangular form, but its third row contains all zeros. There is no way we can bring this into the form of a unit matrix. We therefore conclude that B is not invertible. This is, of course, obvious, since it can easily be verified thatBhas a vanishing determinant. 11II

We mentioned earlier that the determinant is a property of linear lransfonnations, although they are defined in terms of matrices that represent them. We can now show this. First, note that taking the determinant of both sides of AA -1 = I,

3.6 THE TRACE

101

one obtains det(A- 1) = 1/ detA. Now recall that the representations of an operator in two different bases are related via a similarity transformation. Thus, if A is represented by A in one basis and by A' in another, then there exists an invertible matrix R such that A' = RAR- 1 • Taking the determinant of both sides, we get

I det A' = detR detA - - = detA. detR Thus, the determinant is an inttiusic property of the operator, independent of the basis chosen in which to represent the operator.

3.6 The Trace Another inttiusic quantity associated with an operator that is usually defined in terms of matrices is given in the following definition.

trace ota square matrix

3.6.1. Definition. Let A be an N x N matrix. The mapping tr : M N x N -->
tr(AB) = tr(BA).

and

Proof The linearity of the trace and the first identity follow directly from the definition. To prove the second identity of the theorem, we use the definitions of the trace and the matrix product: N

N

N

N

N

tr(AB) = ~)AB)ii = LL(A)ij(B)ji = LL(B)ji(A)ij ;=1

=

f; N

;=1 j=l

~(B)ji(A)ij )

(N

;=1 j=l N = f;(BA)jj = tr(BA).

D connection between trace and determinant

3.6.3. Example. In this example, we show a veryuseful connection between the trace and the determinant that holdswhen a matrix is only infinitesimally different from the unit matrix. Letus calculate thedeterminant of 1 + fA to first order in E. Usingthedefinition of determinant, we write n

det(1 +€A) =

L

£;1 ...iIl(151i1 +€alit)···(l5nin +€a'nin)

i1,.."i n = 1

n

=

E i1,...•in = 1

n

Eit ... in81it···8nin

+€ L

n

E

k=l ;1,...,i n = 1

E"il .. .in81i1···8kik .. ,8ninakik·

102

3. MATRICES: OPERATOR REPRESENTATIONS

The first sumis justtheproduct of all the Kronecker deltas. In the secondsum, 8kik means that in the product of the deltas, ~kik is absent. This term is obtained by multiplying the secondterm of the kth parentheses by the first term of all therest. Since we are interested only in the first power of E, we stop at this term. Now,the first sum is reduced to E12...n = 1 after all the Kronecker deltas are summed over. Forthe second sum,we get n

e

n

n

L E

Eit...i1l8lil··· 8kh ... 8ninetkh

k=l it ....• in=l n

= e

E

=€ L

n

L

€12 ... ik ... nClk i k

k=l ik=l

n

€12 ...k ...netkk = €

k~l

E

akk

= € tr A,

(3.26)

k~l

where thelastlinefollows from thefactthat theonlynonzero value for €12...ik...n isobtained when ik is equalto the missing index, i.e., k, in whichcase it will be 1. ThusOOt(1 + fA) = l+
= tr(RAR- 1) = tr[R(AR- 1)] = tr[(AR-1)R] = tr[A(R-1R)] = tr(A1) = trA.

The preceding discussion is summarized in the following proposition. 3.6.4. Proposition. To every operator A E L (V) are associated two intrinsic numbers, det A and tr A, which are the determinant and trace ofthe matrix representation ofthe operator in any basis of V. It follows from this proposition that the resnlt of Example 3.6.3 can be written in terms of operators:

det(l

+ fA)

=

1+
(3.27)

A par1icnlarly useful formula that can be derived from this equation is the derivative at t = 0 of an operator A(t) depending on a single variable with the property that A(O) = 1. To first order in t, we can write A(t) = 1 + tA(o) where a dot represents differentiating with respect to t. Substituting this in Equation (3.27) and differentiating with respect to t, we obtain the important result -d det(A(t» dt

I

,=0

. = tr A(O).

(3.28)

3.6.5. Example. Wehaveseenthatthedeterminant of a product of matrices is theproduct of the determinants. On the other hand, the trace of a sum of matrices is the sumof traces. When dealing with numbers, products and sums are related via the logarithm andexponential: afJ = exp{lna + In,B}. A generalization of this relation exists for diagonalizable matrices. Let A be sucha matrix, i.e., let D = RAR-1 for-some similarity transformation R and some diagonal matrix D = diag(Al, 1.2, ... , An). The determinantofa diagonalmatris is simplythe product of its elements:

3.7 PROBLEMS

103

Takingthe naturallog of both sides and using the result of Example 3.2.3,we have In(detD) =lnAl +lnA2+···+lnAn =tr(lnD), which can also be written as det D = exp(tr(ln D)). In tenus ofA,thisreadsdet(RAR-1) = exp{tr(ln(RAR- 1))) . Nowinvoketheinvariance of determinant andtrace under similarity transformation andtheresultof Example 3.4.4 to obtain detA = exp{tr(R(lnA)R-1)) = exp{tr(lnA)).

(3.29)

Thisis animportantequation,which is sometimes usedtodefine thedeterminant ofoperators in infinite-dimensional vectorspaces. II

Bolh lhe determinant and lhe trace are mappings from MNxN to C. The determinant is not a linear mapping, bnt lhe trace is; and lhis opens up lhe possibility of defining an inner product in lhe vector space of N x N matrices in terms of lhe trace: 3.6.6. Proposition. For any two matrices A, B E MNxN, the mapping g : MNxN x MNxN --> C defined by g{A, B) = tr(AtB) is a sesquilinear inner product. Proof The proof follows directly from lhe linearity of trace and lhe definition of hermitian conjugate. D

3.7 Problems 3.1. Show lhat if [c) = la) + Ib), lhen in any basis lhe components of Ie) are equal to lhe sums of lhe corresponding components of la) and Ib). Also show lhat lhe elements of lhe matrix representing lhe sum of two operators are lhe sums of the elements of lhe matrices representing IDose two operators. 3.2. Show lhat the unit operator 1 is represented by lhe unit matrix in any basis. 3.3. The linear operator A : ]R3 --> ]R2 is given by A

(X) = (2x + Y - 3Z) . Y x+y-z z

Construct lhe matrix representing A in lhe standard bases of]R3 and]R2. 3.4. The linear transformation T : ]R3 --> ]R3 is defined as

Find lhe matrix representation of T in

104

3. MATRICES: OPERATOR REPRESENTATIONS

(a) the standard basis of JR3,

(b) the basis consisting of lat) = (1,1,0), la2} = (1,0, -I), and la3} = (0,2,3). 3.5. Show that the diagonal elements of an antisymmetric matrix are all zero. 3.6. Show that the nnmber of independent real parameters for an N x N

1. (real) symmetric matrix is N(N + 1)/2,

2. (real) antisymmetric matrix is N(N - 1)/2, 3. (real) orthogonal matrix is N(N - 1)/2, 4. (complex) unitary matrix is N 2 , 5. (complex) hermitian matrix is N 2 • 3.7. Show that an arbitrary orthogonal 2 x 2 matrix can be written in one of the following two forms: COSo ( sinO

- SinO) cos e

or

cos O ( sinO

SinO) -CDSe

.

The first is a pure rotation (its determinant is +I), and the second has determinant -1. The form of the choices is dictated by the assnmption that the first entry of the matrix reduces to I when 0 = O. 3.8. Derive the formulas COS(OI

+~)

= cos 01

cos~

sin(OI +~) = sin 01 cos ~

- sin 01

sin~,

+ cos 01 sin 02

by noting that the rotation of the angle 01 two rotations. (See Problem 3.7.)

+ ~ in the xy-plane is the product of

3.9. Prove that if a matrix Msatisfies MMt = 0, then M = O. Note that in general, M2 = 0 does not imply that Mis zero. Find a nonzero 2 x 2 matrix whose square is zero. 3.10. Construct the matrix representations of and the derivative andmultipiication-by-t operators. Choose II, t, t 2 , t 3 } as your basis of3'~[tl and {l, t , t 2 , t 3 , t 4 } as your basis of3'4[t]. Use the matrix of 0 so obtained to find the first, second, third, fourth, and fifth derivatives of a general polynomial of degree 4.

3.7 PROBLEMS

105

3.11. Find the transformation matrix R that relates the (orthonormal) standard basis of C 3 to the orthonormal basis obtained from the following vectors via the Gram-Schmidt process:

Verify that R is unitary, as expected from Theorem 3.4.2. 3.12. If the matrix representation of an endomorphism T of C 2 with respect to the standard basis is what is its matrix representation with respect to the basis {(D, (!I)}?

(11),

3.I3. If the matrix representation of an endomorphism T of C 3 with respect to the standard basis is

what is the representation of T with respect to the basis

o

{( 1

-I

1

-1



0

),(-1),{

I)}?

3.14. Using Definition 3.5.1, calculate the determinant of a general 3 x 3 matrix and obtain the familiar expansion of such a determinant in terms of the first row of the matrix. 3.15. Prove Corollary 3.5.4. 3.16. Showthatdet(aA) = a.

aN

detAforanNxNmatrixAandacomplexnumber

3.17. Show that det 1 = I for any unit matrix. 3.18. Find a specific pair ofmatrices Aand B such that det(A+ B) oF detA+det B. Therefore, the determinant is not a linear mapping. Hint: Any pair of matrices will most likely work. In fact, the challenge is to find a pair such that det(A + B) = detA + detB. 3.19. Demonstrate Proposition 3.5.5 using an arbitrary 3 x 3 matrix and evaluating the sum explicitly. 3.20. Find the inverse of the matrix A=

(i

-2

-I

o I

106

3. MATRICES: OPERATOR REPRESENTATIONS

3.21. Show explicitly that det(AB) = det Adet B for 2 x 2 matrices. 3.22. Given three N x N matrices A, B, and C such that AB = C with C invertible, show that both A and B must be invertible. Thus, any two operators A and B on a finite-dimensional vector space satisfying AB = 1 are invertible and each is the inverse of the other. Note: This is not true for infinite-dimensional vector spaces. 3.23. Show directly that the similarity transformation induced by R does not change the determinant or the trace of A where

R=

12-1) ( 0

1

-2

2

1 -1

A=

and

3-1 2) ( 0 I

1 -3

-2·. -I

3.24. Find the matrix that transforms the standard basis of 1(;3 to the vectors I

-I

.j2

.j2

j

!al} =

i

!a2) =

.J6

0 !a3) =

.J6

l+i

-1+i

.J6

.J6

-2

.J6 l+i

.J6

Show that this matrix is unitary. 3.25. Consider the three operators LI, L2, and L3 satisfying [Lj , L21 = iL3, [L3, Ll] = iL2. [L2, L3] = iLl. Show that the trace of each of these operators is necessarily zero. 3.26. Show that in the expansion of the determinant given in Definition 3.5.1, uo two elements of the same row or the same column can appear in each term of the sum. 3.27. Find inverses for the following matrices using both methods discussed in this chapter.

A=U -I

D-

I I

-2

CyO 0

1/../2

0

-1) 2 , -2 0

1/../2

2I -1) -2 , 2 1 -1

B= ( 0

(I - i)/(2../2) (I - i)/(2../2)

-(1 - i)/(2../2) 1/../2 -(1 - i)/(2../2) 0

C=

(~1

"H)!('yO)

-(1 -(1

-1

+ i)/(2../2) + i)/(2../2)

I -I

~ ),

-2

J .

(1+ i)/(2../2)

3.28. Let Abe an operator on V. Show that if det A = 0, then there exists a nonzero vector [x) E V such that A [x) = o.

3.7 PROBLEMS

107

3.29. For which values of a are the following matrices iuvertible? Find the iuverses whenever possible.

a A=G

c=

(!

1

a 1

a 0

;), ~),

I

B=G a

1

D=G

I I

a

D, ~) .

3.30. Let {ai}~l' be the set consisting of the N rows of an N x N matrix A and assume that the ai are orthogonal to each other. Show that

Hint: Consider AAt. What would the result be if A were a unitary matrix? 3.31. Prove that a set of n homogeneous linear equations in n unknowns has a nontrivial solution if and only if the determinant of the matrix of coefficients is zero.

3.32. Use determinants to show that an antisymmetric matrix whose dimension is odd cannot have an inverse. 3.33. Show that tr(la} (bl) = (bl a). Hint: Evaluate the trace in an orthonormal basis. 3.34. Show that if two invertible N x N matrices A and B anticommute (that is, AB + BA = 0), then (a) N must be even, and (b) tr A = tr B = O. 3.35. Show that for a spatial rotation Rfi(0) of an angle 0 about an arbitrary axis tr Ril (0) = I + 2 cos O.

n,

3.36. Express the smn of the squares of elements of a matrix as a trace. Show that this smn is invariant under an orthogonal transformation of the matrix. 3.37. Let S and A be a symmetric and an antisymmetric matrix, respectively, and let Mbe a general matrix. Show that (a) trM = tr M', (b) tr(SA) = 0; iu particular, tr A = 0,

(c) SA is antisymmetric if and only if [S, A} = 0, (d) MSM' is symmetric and MAM' is antisymmetric. (e) MHMt is hermitian if His.

108

3. MATRICES: OPERATOR REPRESENTATIONS

3.38. Find the trace of each of the following linear operators: (a) T : ]R3 --> ]R3 given by T(x, y, z) = (x

+Y -

z, 2x

+ 3y -

2z, x - y).

(b) T : ]R3 --> ]R3 given by T(x, y, z) = (y - z, x

+ 2y + z, z -

y).

(c) T : 1(;4 --> 1(;4 given by T(x, y, z, w) = (x

+ iy - z + iw, 2ix + 3y -

2iz - w, x - iy, z + iw).

3.39. Use Equation (3.29) to derive Equation (3.27). 3.40. Suppose that there are two operators A and B such that [A, B] = c1, where c is a constant. Show that the vector space in which such operators are defined cannot be finite-dimensional. Conclude that the position and momentum operators of quantum mechanics can be defined only in infinite dimensions.

Additional Reading 1. Axler, S. Linear Algebra Done Right, Springer-Verlag, 1996. 2. Birkhoff, G. and MacLane, S. Modern Algebra, 4th ed., Macmillan, 1977. Discusses matrices from the standpoint of row and column operations. 3. Greub, W. Linear Algebra, 4th ed., Springer-Verlag, 1975.

4

_

Spectral Decomposition

The last chapter discussed matrix representation of operators. It was pointed out there that such a representation is basis-dependent. In some bases, the operator may "look" quite complicated, while in others it may take a simple form. In a "special" basis, the operator may look the simplest: It may be a diagonal matrix. This chapter investigates conditions under which a basis exists in which the operator is represented by a diagonal matrix.

4.1

Sum oftwo subspaces defined

Direct Sums

Sometimes it is possible, and convenient, to break up a vector space into special (disjoint) subspaces. For instance, it is convenient to decompose the motion of a projectile into its horizontal and vertical components. Similarly, the study of the motion of a particle in R 3 under the influence of a central force field is facilitated by decomposing the motion into its projections onto the direction of angular momentum and onto a plane perpendicular to the angular momentum. This corresponds to decomposing a vector in space into a vector, say, in the xy-plane and a vector along the z-axis. We can generalize this to any vector space, but first some notation: Let 11and W be subspaces of a vector space V. Denote by 11 + W the collection of all vectors in V that can be written as a sum of two vectors, one in 11 and one in W. It is easy to show that 11 + W is a subspace of V. 4.1.1. Example. Let If be the xy-plane and W the yz-plane. These are both subspaces ofJll.3, and so is 1L + W. In fact, 1L + W = Jll.3, because given any vector (x, y, z) inJll.3, we can write it as (x, y, z) = (x,

h. 0) + (0, h. z}.

'-'-' ell

eW

110

4. SPECTRAL DECOMPOSITION

Thisdecomposition is not unique: Wecouldalso write (x, y, z) = (x, anda hostof other relations.

direct sum defined

uniqueness of direct

sum

h, 0) + (0, ~y, z), l1li

4.1.2. Definition. Let U and W be subspaces ofa vector space V such that V = U + Wand the only vector common to both U and W is the zero vector. Then we say that V is the direct sum ofU and Wand write

4.1.3. Proposition. Let U and W be subspaces ofV. Then V = U Ell W if and only if any vector in V can be written uniquely as a vector in 11 plus a vector in W.

Proof Assume V = U Ell W, and let Iv) and a vector in W in two different ways:

Iv) = lu) + Iw) = lu')

+ Iw')

{}

E

V be written as a sum of a vector in U

lu) - lu')

= Iw') -

Iw).

The LHS is in U. Since it is equal to the RHS-which is in W-it must be in W as well. Therefore, the LHS must equal zero, as must the RHS. Thus, lu) = lu') , Iw') = Iw), and there is only one way that [u) can be written as a sum of a vector in U and a vector in W. Conversely, if la) E U and also la) E W, then one can write

-- --

la) = la)

in'lL

+

10)

inW

and

la) = 10)

in'lL

+

la) .

inW

Uniqueness of the decomposition of la) implies that la) = 10). Therefore, the only vector corumon to both U and W is the zero vector. This implies that V = UEilW. D dimensions ina direct sum

4.1.4. Proposition. If V = U Ell W, then dim V = dim U + dim W.

Proof Let llui) }~I be a basis for U and Ilwi)J~=I a basis for W. Then it is easily verified that IluI) ,lu2), ... , lum ) , IWI) ,lw2), ... , IWk)J is a basis for V. The details are left as an exercise. D We can generalize the notion of the direct sum to more than two subspaces. For example, we can write R' = XEIl'IJEIlZ, where X, 'IJ, and Z are the one-dimensional subspaces corresponding to the three axes. Now assume that (4.1) i.e., V is the direct sum of r of its subspaces that have no common vectors among themselves except the zero vector and have the property that any vector in V can be written (uniquely) as a sum of vectors one from each subspace. Define the linear operator Pj by P j lu) = luj) where lu) = LJ=I lu j), luj) E U j . Then it is readily

4.1 DIRECT SUMS

111

verified that P] = P j and P jPk = 0 for j i' k. Thus, the P/s are (not necessarily hermitian) projection operators. Furthermore, for an arbitrary vector lu), we have ¥Iu) E V.

Since this is true for all vectors, we have the identity

(4.2)

orthogonal complement ofa subspace

4.1.5. Definition. Let V be an inner product space. Let JV( be any subspace ofV. Denote by JV(.L the set ofall vectors in V orthogonal to all the vectors in :M. JV(.L (pronounced "em perp") is called the orthogonal complement ofJV(. 4.1.6. Proposition.

JV(.L is a subspace ofV.

Proof In fact, if la), Ib) E JV(.L, then for any vector [c) E JV(, we have

- --=0

(cl (a la)

So, ala)

+ f3lb»

=0

= a (c] a) +f3 (c] b) = O.

+ f3lb} E JV(.L for arbitrary a, f3 E C and la} , Ib) E JV(.L.

o

If V of Equation (4.1) is an inner product space, and the subspaces are mutually orthogonal, then for arbitrary lu) , Iv} E V,

which shows that P j is hermitian. In Chapter 2, we assnmed the projection operators to be hermitian. Now we see that only in an ioner product space (and only if the subspaces of a direct sum are orthogonal) do we recover the hermiticity of projection operators.

4.1.7. Example. Consider an orthonormalbasis BM

= !Iei)):"=l for M, and extend it to a basis B = !Iei))~l for V. Now coostruct a (hermitian) projection operator P = L~l lei) {ei I· Thisis the operatorthat projects anarbitrary vector in Vanta the subspace M. It is straightforward to showthat 1 - P is the projectionoperatorthat projectsonto M.L (see Problem4.2). An arbitrary vector la) E V canbe written as

la)

= (P + 1 -

P) la)

= Pia) + (1 -

P) la) .

'-.-'

in .M:

in :M.l

Furthermore, theonlyvector that canbe in both:M andMl..is thezerovector, because it is the only vectororthogonalto itself. III

112

4. SPECTRAL DECOMPOSITION

From this example and the remarks immediately preceding it we may conclude the following: 4.1.8. Proposition. If V is an inner product space, then V = M Ell M.L for any subspace M. Furthermore, the projection operators corresponding to M and M.L are hermitian.

4.2

Invariant Subspaces

This section explores the possibility of obtaioiog subspaces by means ofthe action of a linear operator on vectors of an N-dimensional vector space V. Let 10) be any vector 10V, and A a linear operator on V. The vectors 10) ,A 10) ,A2 1a) , ... , AN 10)

are linearly dependent (there are N + I of them!). Let M es Span {Ak 10) Jf=o' It follows that, m es dim M =" dim V, and M has the property that for any vector [x) E M the vector A [x) also belongs to M (show this!). In other words, no vector 10 M "leaves" the subspace wheu acted on by A. invariant subspace: reduction ofan operator

4.2.1. Definition. A subspace M is an invariant subspace ofthe operator A if A transforms vectors ofM into vectors ofM. This is written succinctly as A(M) C M. We say that M reduces A if both M and M.L are invariant subspaces of A. Startiog with a basis of M, we can extend it to a basis B = {Iai) J!:, of V whose first m vectors span M. The matrix representation of A 10 such a basis is given by the relation Alai) = I:f=''''jilaj),i = 1,2, ... ,N.lfi =" m,then ",ji = 0 for j > m, because Alai) belongs to M when i =" m and therefore can be written as a linear combioation of only {Ia,) , 102), ... , lam)). Thus, the matrix representation of A 10 B will have the form A = (An 02'

matrix representation ofan operator in a subspace

block diagonal matrix defined

where An is anm x m matrix, A'2 anm x (N - m) matrix, 021 the (N - m) x m zero matrix, and A22 an (N - m) x (N - m) matrix. We say that An represents the operator A 10the m-dimensional subspace M. It may also happen that the snbspace spanned by the remaioiog basis vectors 10 B, namely lam+, ) , lam+2) , ... , ION), is also an invarianl subspace of A. Then A12 will be zero, and A will take a block diagonal form:' A = (An o

reducible and irreducible matrices

A12), A22

0) A22

A matrix representation of an operator that can be brought ioto this form by a suitable choice of basis is said 10 be reducible; otherwise, it is called irreducible.

1Fromnow on, we shalldenoteall zeromatrices by the same symbolregardless of theirdimensionality.

4,2 INVARIANT SUBSPACES

113

A reducible matrix A is denoted in two different waysr' (4.3)

For example, when M reduces A and one chooses a basis the first m vectors of which are in M and the remaining ones in M 1-, then A is reducible. We have seen on a number of occasions the significance of the hermitian conjugate of an operator (e.g., in relation to hermitian and unitary operators). The importance of this operator will be borne out further when we study the spectral theorem later in this chapter. Let us now investigate some properties of the adjoint of an operator in the context of invariant subspaces. condition for invariance

4.2.2. Lemma. A subspace M ofan inner product space V is invariant under the linear operator A ifand only ifM1- is invariant under At,

Proof The proof is left as a problem.

D

An immediate consequence ofthe above lemma and the two identities (At) t = A and (M1-)1- = M is contained in the following theorem. 4.2.3. Theorem. A subspace of V reduces A if and only if it is invariant under both A and At. 4.2.4. Lemma. Let M be a subspace ofV and P the hermitian projection operator onto M. Then M is invariant under the linear operator A ifand only if AP = PAP,

Proof Suppose M is invariant. Then for any [x) in V, we have Pix) EM·=} AP Ix) EM=} PAP Ix) = AP Ix) .

Since the last equality holds for arbitrary [x), we bave AP = PAP, Conversely, suppose AP = PAP. For any IY) E M, we have Ply)

= IY)

=}

AP Iy)

'-.-'

= A Iy) = P(AP IY»

E M.

=PAP

Therefore, M is invariant under A.

D

4.2.5. Theorem. Let M be a subspace of V, P the hermitian projection operator of V onto M, and Aa linear operator on V. Then M reduces A ifand only if A and P commute.

21tis common to use a single subscript for submatrices of a block diagonal matrix, just as it is common to use a single subscript for entries of a diagonal matrix.

114

4. SPECTRAL DECOMPOSITION

Proof Suppose M reduces A. Then by Theorem 4.2.3, M is invariant under both A and At. Lemma 4.2.4 then implies AP = PAP

and

Atp

= PAtp.

(4.4)

Taking the adjoint of the second equation yields (Atp)t = (PAtp)t, or PA = PAP. This equation together with the first equation of (4.4) yields PA = AP. Conversely, suppose that PA = AP. Then p 2A = PAP, whence PA = PAP. Taking adjoints gives Atp = PAtp, because P is hermitian. By Lemma 4.2.4, M is invariant under At. Similarly, from PA = AP, we get PAP = Ap2 , whence PAP = AP. Once again by Lemma 4.2.4, M is invariant under A. By Theorem 4.2.3, M rednces A. D The main goal of the remaining part of this chapter is to prove that certain operators, e.g. hermitian operators, are diagonalizable, that is, that we can always find an (orthonormal) basis in which they are represented by a diagonal matrix.

4.3 Eigenvalues and Eigenvectors Let us begin by considering eigenvalues and eigenvectors, which are generalizations of familiar concepts in two and three dimensions. Consider the operation of rotation about the z-axis by an angle 0 denoted by Rz(O). Such a rotation takes any vector (x, y) in the xy-plane to a new vector (x cos 0 - y sin 0, x sin 0 + y cos 0). Thus, unless (x, y) = (0,0) or 0 is an integer multiple of 2n, the vector will change. Is there a nonzero vector that is so special (eigen,in German) that it does not change when acted on by Rz(O)? As long as we confine ourselves to two dimensions, the answer is no. But if we lift ourselves up from the two-dimensional xy-plane, we encounter many such vectors, alI of which lie along the z-axis. The foregoing example can be generalized to any rotation (normally specified by EuIer angles). In fact, the methods developed in this section can be used to show that a general rotation, given by Euler angles, always has an unchanged vector lying along the axis around which the rotation takes place. This concept is further generalized in the following definition. eigenvalue and eigenvector

4.3.1. Definition. A scalar A is an eigenvalue and a nonzero vector la} is an eigenvector of the linear transformation A E J:., (V) if Ala} = Ala).

(4.5)

4.3.2. Proposition. Add the zero vector to the set ofall eigenvectors ofAbelonging to the same eigenvalue A, and denote the span of the resulting set by M A . Then M A is a subspace of V, and every (nonzero) vector in M A is an eigenvector of A with eigenvalue A. Proof The proof follows immediately from the above definition and the definition of a subspace. D

4.3 EIGENVALUES AND EIGENVECTORS

eigenspace

spectrum

115

4.3.3. Definition. The subspace JVe,. is referred to as the eigenspace ofA corresponding to the eigenvalue ),.. Its dimension is called the geometric multiplicity of ),.. An eigenvalue is called simple if its geometric multiplicity is 1. The set of eigenvalues ofA is called the spectrum ofA. By their very construction, eigenspaces corresponding to different eigenvalues have no vectors in common except the zero vector. This can be demonstrated by noting that if Iv) E. M, n MIL for ), f= 1-'-, then 0= (A - ),.1) Iv) = A Iv) - ),.Iv) = I-'-Iv) - ),.Iv) = (I-'- -),.) Iv)

,.0

'--.-'

'* Iv) = O.

Let us rewrite Equation (4.5) as (A - ),,1) la) = O. This equation says that la) is an eigenvector of A if and only if la) belongs to the kernel of A - ),,1. If the latter is invertible, then its kernel will consist of only the zero vector, which is not acceptable as a solution of Equation (4.5). Thus, if we are to obtain nontrivial solutions, A -),,1 must have no inverse. This is true if and only if det(A - ),,1) = O.

characteristic polynomial and characteristic roots ofan operator

(4.6)

The determinant in Equation (4.6) is a polynomial in)", called the characteristic polynomial of A. The roots of this polynomial are called characteristic roots and are simply the eigenvalues of A. Now, any polynomial of degree greater than or equal to I has at least one (complex) root, which yields the following theorem. 4.3.4. Theorem. Every operator on a finite-dimensional vector space over iC has at least one eigenvalue and therefore at least one eigenvector. Let X}, ),.2, •.• ,),.p be the distinct roots of the characteristic polynomial of A, and let x j occur m j times. Then 3 p

det(A - ),,1) = (),.I -

),.)m 1 ••• (),.p -

),.)mp =

D

(),.j _ ),.)mj.

(4.7)

j~1

For),. = 0, this gives p 'l.rn t , m 2 'l.mp_D'l.mj detA -- 11.1 11.2 ••• "-p Aj.

(4.8)

j=l

Equation (4.8) states that the determinant of an operator is the product of all its eigenvalues. In particular, if one of the eigenvalues is zero, then the operator is not invertible. 3 m j is called the algebraic multiplicity of J.. i-

116

4. SPECTRAL DECDMPDSITIDN

4.3.5. Example. Let us fiud the eigenvaluesof a projectionoperator P.1f la} is an eigenvector,then Pia) = Ala}. Applying P on both sides again, we obtain

= AP la} = 1.(1. la}) = Alia). But p2 = P; thus, Pia} = Alia). It follows thaUlla) = A[e},or (1.1 p2la)

la}

i' 0, wemustbaveA(A-I)

=

A) la} O. Since = 0, or 1.= 0, 1. Thus, the only eigenvalues of a projection

operator are 0 and 1. Thepresence of zero as an eigenvalue of P is an indication that P is not invertible. II 4.3.6. Example. Tobe able to see thedifference betweenalgebraic andgeometric multiplicities,considerthe matrixA = (A 1), whose characteristicpolynomialis (I - 1.)1. Thus, the matrix has only one eigenvalue, A = I, with algebraic multiplicity ml = 2. However, the most generalvector la) satisfying(A-1) la) = 0 is easily shown to be ofthefonn (~). This showsthat M).=l is one-dimensional, i.e., the geometric multiplicity of A is 1. l1li

As mentioned at the beginning of this chapter, it is useful to represent an operator by as simple a matrix as possible. The simplest matrix is a diagoual matrix. This motivates the following definition: diagonalizable operators

4.3.7. Definition. A linear operator A on a vector space V is said to be diagonalizable if there is a basisfor V all ofwhose vectors are eigenvectors of A. 4.3.8. Theorem. Let A be a diagonalizable operator on a vector space V with distinct eigenvalues {Aj }j~t. Thereare (not necessarily hermitian)projection operators Pj on V such that r

(I) 1 = LPj,

r

(2) PiPj = 0 for i

t= i.

j=l

(3) A = I>jPj. j=l

Proof Let JY(j denote the eigenspace corresponding to the eigenvalue Ai- Since the eigenvectors span V and the only common vector of two different eigenspaces is the zero vector (see comments after Definition 4.3.3), we have

This immediately gives (I) and (2) if we use Equations (4.1) and (4.2) where Pj is the projection operator onto JY(sTo prove (3), let Iv} be an arbitrary vector in V. Then [u) can be written uniquely as a sum of vectors each coming from one eigenspace. Therefore,

A [u)

= ~A IVj) = ~Aj IVj) = (~AjPj) Iv}.

Since this equality holds for all vectors Iv}, (3) follows.

D

4.4 SPECTRAL DECOMPOSITION

4.4

117

Spectral Decomposition

This section derives one of the most powerful theorems in the theory of linear operators, the spectral decomposition theorem. We shall derive the theorem for operators that generalize hermitian and unitary operators. normal operator defined

4.4.1. Definition. A normal operator is an operator on an inner product space that commutes with its adjoint.

An important consequence of this definition is that if and only if

Ais normal.

(4.9)

4.4.2. Proposition. Let A be a normal operator on V. Then Ix} is an eigenvector of A with eigenvalue). ifand only iflx} is an eigenvector of At with eigenvalue): *. Proof By Equation (4.9), thefactthat(A-)"1)t = At -),,*1, and the factthatA-)"1 isnonnal (reader, verify), we have II (A-)"1)xll = 0 ifandonly if II (At -)..*1)xll = O. Since it is only the zero vector that has the zero nonn, we get (A - ),,1) Ix} = 0

if and only if

(At - ),,*1) [x) =

o. D

This proves the proposition.

We obtain a useful consequence of this proposition by applying it to a hermitian operator H and a unitary operator" U. In the first case, we get

Therefore, Ais real. In the second case, we write Ix)

= 1jx) = UUt

[x)

= U()..* Ix)) = )..*U Ix) = U* Ix) =>

U*

= I.

Therefore, ).. is unimodular (has absolute value equal to 1). We summarize the foregoing discussion: 4.4,3. Corollary. The eigenvalues ofa hermitian operator are real. The eigenvalues ofa unitary operator have unit absolute value.

4.4.4. Example. Let us find the eigenvalues and eigenvectors of the hermitian matrix H = (~r})' We have -A

det(H -Al) = det ( i

Thus, theeigenvalues, Al = 1 andA2 "Obviously, bothare normal operators.

= -I, arereal, as expected.

118

4. SPECTRAL DECOMPOSITION

Tofindthe eigenvectors, we write

-i) ("I) = (--:"1 - i"2) cz la'i - cz

-1

oraz = ial. which gives lal) = (iC:X1r ) = Atso,

al (}), where cq

is an arbitrary complex number.

-i) (Ih) = (fh 1

Always normalize the eigenvectors!

fh

ifh) zfh + fh

or fh = -ilh, which gives la2} = (!'~1) = PI <-~;l, where PI is an arbitrary complex number. It is desirable, in mostsituations, to orthonormalize theeigenvectors. In thepresent case, theyarealready orthogonal. Thisis a property shared by all eigenvectors of a hermitian (in fact,normal) operator stated in the next theorem. We therefore need to merelynormalize the eigenvectors:

or 1"11 = 1/.J2and "I = ei~ 1.J2 for some 'P E lll.. A similarresull is obtained for PI. The choice ip = 0 yields

and

III

The following theorem proves for all normal operators the orthogonality property of their eigenvectors illnstrated in the example above for a simple hermitian operator.

4.4.5. Theorem. An eigenspace ofa normal operator reduces that operator. Moreover, eigenspaces of a normal operator are mutually orthogonal. Proof The first part of the theorem is a trivial consequence of Proposition 4.4.2 and Theorem 4.2.3. To prove the second part, let Iu) E JY[A and Iv) E JY[I' with A i' 1-'- Then, using Theorem 4.2.3 once more, we obtain A(vlu) = (vIAu)

= (vIAu) = (Atvlu) = (Il*vlu) =Il(vlu)

It follows that (A - Il) (vi u) = 0 and since Ai' u, (vi u) = O.

spectral decomposition theorem

D

4.4.6. Theorem. (Spectral Decomposition Theorem) Let A be a normal operator on afinite-dimensional complex inner product space V. Let AI, A2, ... , A, be its distinct eigenvalues. Then there exist nonzero (hermitian) projection operators PI, Pz, ... , Pr such that

Vi i' j ..

4.4 SPECTRAL DECDMPDSITIDN

119

3. L;'=I AiPi = A.

Proof Let Pi be the operator that projects onto the eigenspace Mi corresponding to eigenvalue Ai. By conuuents after Proposition 4.1.6, these projection operators are hermitian. Because of Theorem 4.4.5, the only vector conuuon to any two distinct eigenspaces is the zero vector. So, it makes sense to talk about the direct sum of these eigenspaces. Let M = MI E9M2E9· . 'E9Mr and P = Lr=1 Pi, where P is the orthogonal projection operator onto M. Since A conuuutes with every Pi (Theorem 4.2.5), it conuuutes with P. Hence, by Theorem 4.2.5, M reduces A, i.e., M 1- is also invariant under A. Now regard the resttiction of A to M 1- as an operator in its own right on the finite-dimensional vector space M 1-. Theorem 4.3.4 now forces A to have at least one eigenvector in M 1-. But this is impossible because all eigenvectors of A have been accounted for in its eigenspaces. The only resolution is for M 1- to be zero. This gives and The second equation follows from the first and Equations (4.1) and (4.2). The remaining part of the theorem follows from arguments sintilar to those used in the D proof of Theorem 4.3.8. We can now establish the counection between the diagonalizability of a normal operator and the spectral theorem. In each subspace Mi' we choose an orthonormal basis. The union of all these bases is clearly a basis for the whole space V. Let us label these basis vectors lei), where the subscript indicates the subspace and the superscript indicates the particnlar vector in that subspace. Clearly, (ei lei:) =

= L;~llei){eil. Noting that Pk the mattix elements of A in such a basis:

8",8 jj , and Pj

len = 8kj' lei:), we can obtain

Only the diagonal elements are nonzero. We note that for each subscript j we have

m j orthonormal vectors lei}, where m j is the dimension of M i- Thus, Aj occurs m j times as a diagonal element. Therefore, in such an orthonormal basis, A will be represented by ... , AI, A2, ... , A2,. ..., , Ar , ••• , Ar.) . diag(AI, '"-.,,-" ml times

m2

times

m- times

Let us sunuuarize the foregoing discussion: 4.4.7. Corollary. If A E £'(V) is normal, then V has an orthonormalbasis consist-

ing ofeigenvectorsof A. Therefore, a normal operatoron a complex innerproduct space is diagonalizable.

120

4. SPECTRAL DECOMPOSITION

Using this corollary, the reader may show the following:

4.4.8. Corollary. A hermitianoperator is positive if and only if all its eigenvalues are positive.

4.4.9. Example. Computation of the largest and the smallest eigenvalues ofa normal operator

COMPUTATION OF LARGEST AND SMALLEST EIGENVALUES

There is an elegant technique that yields the largest andthe smallest (in absolute value) eigenvalues of a normal operator A in a straightforward way if the eigenspaces of these eigenvalues are onedimensional. Forconvenience, assume that theeigenvalues are labeled in order of decreasing absolute values:

IAII > IA21 > ... > IArl ;6 O. Let (laknt'=1 be a basis of V consisting of eigenvectorsof A, and Ix) = Lt'=1 ~k lak) an arbitrary vectorin V. Then Am Ix) =

f ~kAm

lak) =

k=1

f ~kAT

lak) = A'i'

k=1

i:

[~llal) + ~k k=2

e

flak)] .

I

In the limit m --+ 00, the summation in thebrackets vanishes. Therefore,

and for any Iy} obtain

E

V. Taking theratio of this equation and the corresponding one for m + 1, we

. (yl Am + 1 Ix) lim m-e-co (yl Am Ix)

AI.

Notehow crucially thisrelation depends on thefactthat Al is nondegenerate, i.e., that JV(1 is one-dimensional. By taking larger andlarger values for m, we can obtain a better and better approximation to thelargest eigenvalue. Assuming that zero is notthesmallest eigenvalue Ar-and therefore notaneigenvalueof A, we can find the smallesteigenvalueby reptacing A with A-I and Al with liAr. The details areleft as anexercise forthereader. III Ahermitian matrix can be diagonalized bya unitary matrix.

Any given hermitian matrix H can be thought of as the representation of a hermitian operator in the standard orthonormal basis. We can find a unitary matrix U that Can transform the standard basis to the orthonormal basis consisting of lej), the eigenvectors of the hermitian operator. The representation of the hermitian operator in the new basis is UHUt, as discussed in Section 3.3. However, the above argument showed that the new matrix is diagonal. We therefore have the following result.

4.4.10. Corollary. A hermitianmatrix can always be broughtto diagonalform by meansofa unitarytransformation matrix.

4.4.11. Example. Letus consider thediagonalization of thehermitian matrix

o

o -I-i 1- i

-I+i 1+ i -I+i -I-i)

o o

o

o

.

4.4 SPECTRAL DECOMPOSITION

121

The characteristic polynomial is det(H - A1) = (A + 2)2(A - 2)2 Thus, Al = -2 with

multiplicity m 1 = 2, and).2 = 2 With multiplicity m2 = 2. To findthe eigenvectors, we first look at the matrix equation (H + 21) la} = 0, or

o 2 -1-;

1- i

-I-.i) GI)

-1+; -1+;

I+!

o

2

o

"'2

=0.

"'3

2

4

Thisis a systemof linear equations whose "solution" is

Wehavetwo arbitrary parameters, so we expecttwo linearlyindependent solutions. Forthe two choices al

= 2, cz = 0 and al = 0, Cl2 = 2, we obtain, respectively,

and

which happen to he orthogonal. We simply normalize them to ohtain

and

"'3

Similarly, the second eigenvalue equation, (H- 21) la} = 0, gives rise to the conditions "'2), whichprodnce the orthonormal

= -!(l + ;)("'1 +"'2) and "'4 = -!(l- ;)("'1 -

vectors

and

I", (

le4} =

2...,2

-

I~

.).

-!

1-; The unitary matrix thatdiagonaIizes Hcan be constructed from these colunmvectors using the remarks hefore Example 3.4.4, which imply that if we simply put the vectors lei} together as columns, theresulting matrix is Ut :

Ut = _

I

2-,/2

2

0

2

0

0

2

0

2

1+;

1+;

-1-; -1-;

1-; -1+; -1+;

1- i

122

4. SPECTRAL DECOMPOSITION

and the unitary matrix will be

2

0

1- i

1+ i

0

2

1- i

-1- i

2,f'i 2 0

-I+i

-1- i

2

-I+i

1+ i

1

U=(Ut)t=_,

0

We can easily check that U diagonalizes H, i.e., that UHU t is diagonal.

application ot diagonalizalion in electromagnetism

4.4.12. Example. In some physical applications the ability to diagonaIize matricescan be very useful. As a simple but illustrative example, let us consider the motion of a charged particle in a constant magnetic field pointing in the z direction. The equation of motion for such a particle is

d m--.! = qv x B = qdet dt

ey (eXVx Vy 0

o

which in component form becomes

dv y dt

qB m

- = - - Vx ,

dv z dt

-=0.

Ignoring the uniform motion in the z direction, we need to solve the first two coupled equations, which in matrix form becomes d dt

(Vv x ) = y

qB ( 0

(4.10)

-;;; -1

where we have introduced a factor of i to render the matrix hermitian, and defined (j) = qB/m. If the 2 x 2 matrix were diagonal, we would get two uncoupled equations, which we could solve easily. Diagonalizing the matrix. involves finding a matrix R such that

If we could do such a diagonalization, we wouldmuttipty (4.10) by Rto get'

which can be written as

d -d t

(vi) _ . (1'1 ,-

vy

-leV

0

where

5The fact that R is independent of t is crucial in this step. This fact, in tum, is a consequence of the independence from t of the original 2 x 2 matrix.

4.4

SPECTRAL DECDMPDSITIDN

123

Wethenwouldhavea pair of uncoupled equations

that have v'XOx = v' e-iwtttt and v'Y=Ov'y e-iw!LZ t as a solutionset'inO which v' and v'Oy x are integration constants. To find R, we need the normalized eigenvectors of (~i But these are obtained in preciselythe samefashion as in Example 4.4.4. Thereis, however, an arbitrariness in the

b).

solutions due to the choice in numbering the eigenvalues. If we choose the normalized

eigenvectors le2)

1 (-i) = -J'i 1 '

then from comments at the end of Section 3.3, we get

-1

R

t

(i

1

t t = 1R=(R)

=R=-J'il

(-i.

-J'i'

With this choice of R,we have

so that 1'1

= 1 = -1'2. Havingfound Rt, we can write

x) = (V vy

Rt

(VvJ) = y

_1

-J'i

(Vbxe~i"") (i1 -i) 1 v' e l wf

Oy

.

(4.11)

If the x and y components of velocity at t = 0 are VOx and VDy' respectively, then

= Rt (V 9X) , (vvOox) vOy y

or

9X) = R (vox) = _1 (-iVo x+ VOY) . (V vOy VDy..,fi i vOx + VO y

Substituting iu (4.11), we obtain

-i) «-iVox + voy)e~iwt) = 1

(ivox+voy)elwt

( vOx cos cr + VDy sin cr ) -vOxsinwt+voyCOSW( .

This gives the velocity asa function of time. Antidifferentiating once with respect to time yields the position vector. II

In many situations of physical interest, it is desirable to know whether two operators are simultaneously diagonalizable. For instance, if there exists a basis of a Hilbert space of a quantum-mechanical system consisting of simultaneous eigenvectors of two operators, then one can measure those two operators at the same time. In particular, they are not restricted by an uncertainty relation.

simultaneous diagonalization defined

4.4.13. Definition. Two operators aresaidto be simultaneouslydiagonalizable if they can be written in terms a/the same set a/projection operators, as in Theorem 4.4.6.

124

4. SPECTRAL OECOMPOSITION

This definition is consistent with the matrix representation ofthe two operators, because if we take the orthonormal basis B = (Iej) ) discussed right after Theorem 4.4.6, we obtain diagonal matrices for both operators. What are the conditions under which two operators can be simultaneously diagonalized? Clearly, a necessary

condition is that thetwo operators commute. This is animmediate consequence of the orthogonality of the projection operators, which trivially implies Pi Pj = PjPi for all i and j. It is also apparent in the matrix representation of the operators: Any two diagonal matrices commute. What about sufficiency? Is the commutativity of the two operators sufficient for them to be simultaneously diagonalizable? To answer this question, we need the following lemma: 4.4.14. Lemma. An operator T commutes with a normal operator A if and only

if

T commutes with all the projection operators of A.

Proof The "if" part is trivial. To prove the "only if" part, suppose AT = TA, and let Ix} be any vector in one of the eigenspaces of A, say M i- Then we have A(T Ix}) = T(A Ix}) = T(A.j Ix}) = Aj(T Ix»); i.e., T Ix) isinMj,orMj is invariant under T. Since M j is arbitrary, T leaves all eigenspaces invariant. In particular, it leaves M the orthogonal complement of M j (the direct sum of all the remaining eigenspaces), invariant. By Theorems 4.2.3 and 4.2.5, TPj = PjT; and this holds forallj. 0

t,

necessary and sufficient condition forsimultaneous diagonalizability

4.4.15. Theorem. A necessary and sufficient condition for two normal operators A and B to be simultaneously diagonalizable is [A, B] = O.

Proof As claimed above, the "necessity" is trivial. Toprovethe "sufficiency," let A = Lj~l AjPj and B = Lf~l I-'kQk, where {Aj} and {Pj} are eigenvalues and projections of A, and {M} and {Qk! are those of B. Assume [A, B] = O. Then by Lemma 4.4.14, AQk = QkA. Since Qk commutes with A, it must commute with the latter's projection operators: PjQk = QkPi- To complete the proof, define Rjk sa PjQk, and note that

= (PjQk)t = Qlpj = QkPj = PjQk = Rjk, (Rjk)2 = (PjQd = PjQkPjQk = PjPjQkQk = PjQk = Rjk. Rjk

Therefore, Rjk are hermitian projection operators. Furthermore, r

r

LRjk=LPjQk=Qk. j=l

j=l

Similarly, and B as

=, ----Lf=l = Lf=l

r

A = LAjPj j=l

Rjk

r

P jQk

s

= LLAjRjk, j~l k~l

= Pj Lf=l Qk = Pi- We can now write A ,

,

r

B = LI-'kQk = LLI-'kRjk. k~l

k~lj=l

125

4.5 FUNCTIONS OF OPERATORS

By definition, they are simultaneously diagonalizable. spectral decomposition ofa Pauli spin matrix

D

4.4.16. Example. Let us findthe spectraldecompositionof the Paoti spin matrix 0

U2 = ( i

-i) 0 .

The eigenvalues andeigenvectors have been found in Example 4.4.4. These are

and The subspaces JY[J.. j areone-dimensional; therefore,

PI

1 (1)i -J'i1 (1 = 1eJ) (ell = -J'i

P2 = !e2) (e21

= -21(1) -z. (l

i)

-i)

1(1i =:2

-i) 1 '

= -21(1- I.

(6?) and -i) _ ~ (1

Wecan check that PI + P2 =

1

III

2-'

What restrictions are to be imposed on the most general operator T to make it diagonalizable? We saw in Chapter 2 that T can be written in terms of its socalled Cartesian components as T = H + i H' where both H and H' are hermitian and can therefore be decomposed according to Theorem 4.4.6. Can we conclude that T is also decomposable? No. Because the projection operators used in the decomposition of H may not be the same as those used for H'. However, if H and H' are simultaneously diagonalizable such that r

and

H' =

I>~Pko

(4.12)

k~1

then T = L:k~1 ()'k + iA~)Pk. It follows that T has a spectral decomposition, and therefore is diagonalizable. Theorem 4.4.15 now implies that H and H' must commute. Since, H = ~ (T + Tt) and H' = (T- Tt), we have [H, H'] = 0 if and ouly if [T, Tt] = 0; i.e., T is normal. We thus get back to the condition with which we started the whole discussion of spectral decomposition in this section.

tr

4.5 Functions of Operators Functions of transformations were discussed in Chapter 2. With the power of spectral decomposition at our disposal, we can draw many important conclusions about them.

126

4. SPECTRAL OECOMPOSITION

First, we note thatifT = L;~l A,P" then, because of orthogonality of the P,'s

, 1

2

, LAfPi.

= LATPi"'" yn = i=1

;=1

Thus, any polynomial pinT has a spectral decomposition given by p(T) = L;=l p (Ai) Pi . Generalizing this to functions expandable in power series gives 00

(4.13)

f(T) = Lf(Ai)Pi. i=l

4.5.1. Example. Let us investigate the spectral decomposition of the following unitary (actually orthogonal) matrix:

-Sine) cos a . Wefindthe eigenvalues

det (

-SinB)

CaSe - A . 0

0' A

cos\] -

SIn v

z =1. -2cosOA+I=0.

yieldingAl = e- i8 andA2 = ei9 . ForAl we have (reader, provide the missingsteps) e i () sinO

COS B (

CO~9S~~il~) (:~)= 0 ~

a2

=>

= ial

and for A2. COS (

e- e-i()

- sinO ) ("I) 0 cosB - e-i 9 a2 =

sinB

=>

"z = -i"l

=>

lez) =

Jz CJ .

We note that the Ml j are one-dimensionaland spanned by [ej). Thus, PI = leI) (ell = Pz = lez) (ezl

~ (;)(1

-i) =

2:I(Ii

-i) I '

=!2 (I.) (I O=!(l -, 2 -,



Clearly, PI + Pz = 1, and e

-i8

i9

PI + e Pz =

1

2

(e-i() ie-i8

_ie-iO) e-te

I ( eiO . te -Ie

+ 2-

iO) ieif) = U. e

If we takethe natural log of this equation anduse Equation (4.13), we obtain

In U = In(e-iO)Pl + In(eiO)pz = -iOPI + ieP z

= i( -OPI + OPz) == iH,

(4.14)

where H == -fjPl +8P2 is ahennitianoperatorbecausefj is realandPI andP2are hermitian. InvertingEquation (4.14) gives U = ei " , where H = O(-PI

+ Pz) = 0 (~i ~).

III

4.5 FUNCTIONS OF OPERATORS

127

The example above shows that the unitary 2 x 2 matrix U can be written as an exponential of an anti-hermitianoperator. This is a general result. In fact, we have the following theorem, whose proof is left as an exercisefor the reader (see Problem 4.22).

4.5.2. Theorem. A unitary operator U on afinite-dimensional complex inner productspace canbe writtenas U = ei H where H is hermitian. Furthermore, a unitary matrix can be brought to diagonal form by a unitary transformation matrix.

The last statemeutfollows from Corollary 4.4.10 and the fact that

The square root ofan operator isplagued by multivaluedness. In the real numbers, we have only two-valued ness!

for any functiou f that can be expanded in a Taylorseries. A useful function of an operator is its square root. A natural way to definethe square root of an operator A is .jA = L~~l (±A)Pi. This clearly gives many candidatesfor the root, because each term in the sum can have either the plus sign or the mious sign.

4.5.3. Definition. The positive squareroot ofa positive operator A = is.;A = L~~l APi.

L~=l A,P,

The uniquenessof the spectral decompositionimplies that the positive square root of a positive operatoris unique.

4.5.4. Example. Letus evaluate ,fA where

A=CS3i

~).

First,we have to spectrally decompose A.Its characteristic equation is 1.2 - lOA + 16 = 0,

withroots Al = 8 andA2 = 2. Since both eigenvaluesare positive andA is hermitian, we conclude that Ais indeedpositive (Corollary 4.4.8). We can also easily find its normalized eigenvectors:

and

1<2)

1

=.,fi

(-i) 1 .

Thus,

;)

-i) 1 '

,

and

-i) 1 (3 1 =.,fi -i ~) . We caneasily check !hat (,fA)2

=

A.

III

128

4. SPECTRAL DECOMPOSITION

Intuitively, higher and higher powers of T, when acting on a few vectors of the space, eventually exhaust all vectors, and further increase in power will be a repetition of lower powers. This intuitive idea can be made more precise by looking at the projection operators. We have already seen that Tn = LJ=l AjPj, n = 1,2, .... For various n's one can "solve" for Pj in terms of powers ofT. Since there are only a finite number of P j 's, only a finite number of powers of T will suffice. In fact, we canexplicitly construct the polynomial in T for Pi- If there is such a polynomial, by Equation (4.13) it must satisfy Pj = pj(T) = L~~l Pj(Ak)Pk, where Pj is some polynomial to be determined. By orthogonality of the projection operators, Pj O.k) must be zero unless k = j, in which case it mnst be I. In other words, P j O.k) = 8kj. Such a polynomial can be explicitly constructed:

Therefore, (4.15)

and we have the following result.

4.5.5. Proposition. Every function ofa normal operator on a finite-dimensional vector space can be expressed as a polynomial. In fact.from Equations (4./3) and (4.15), (4.16)

4.5.6. Example. Let us write .,fA of the last example as a polynomial in A. We have

Substituting in Equation (4.16), we obtain

.,fA = j):j PI (A) +../i:i.P2(A) =

v'8

.,fi

.,fi

v'8

6

6

6

3

-(A -2) - - ( A - 8) = - A + - .

The RHS is clearly a (first-degree) polynomial in A, and it is easy to verify that it is the matrix of ..fA obtainedin the previousexample. II

4.6 POLAR DECOMPOSITION

129

4.6 Polar Decomposition We have seen many similarities between operators and complex numbers. For instance, hermitian operators behave very much like the real numbers: they have

realeigenvalues; their squares arepositive; everyoperator canbe written as H+i H', where both H and H' are hermitian; and so forth. Also, unitary operators can be written as expiH, where H is hermitian. So unitary operators are the analogue of complex numbers of unit magnitude such as eiO• A general complex number can be written as re'", Can we write an arbitrary operator in an analogous way? The following theorem provides the answer. polar decomposition theorem

4,6.1. Theorem. (polar decomposition theorem) An operator A on afinite-dimensionalcomplexinnerproductspace can be writtenas A = UR where R is a (unique)

positive operatorand U a unitary operator. If A is invertible,then U is also unique. Proof We will prove the theorem for the case where the operator is invertible. The proof of the general case can be found in books on linear algebra (such as [Halm 58]). The reader may show that the operator At A is positive. Therefore, it has a unique positive square root R. We let V = RA-I, or VA = R. Then

w t = RA-1(RA-1)t = RA-1(A-1)tRt = R(AtA)-IRt

= R(R2)-I Rt = R(RtR)-IRt = RR-1(Rtl-1Rt = 1 and

and V is indeed a unitary operator. Now choose U = vt to get the desired decomposition. U'R' implies that R UtU'R' and To prove uniqueness we note that UR R2 = RtR = (utu'R'lt(utU'R') = R'tU'tUUtU'R' = R'tR' = R'2. Since the positive transformation R2 (or R'2l has only one positive square root, it follows that R = R' If A is invertible, then so is R = UtA. Therefore,

=

UR

= U'R

=}

URR- 1 = U'RR- 1

=}

=

U

= U', D

and U is also unique.

It is interesting to note that the positive definiteness of R and the nonuniqueness of U are the analogue of the positivity of r and the nonuniqueness of eiO in the polar representation of complex numbers:

z = re i fJ

=

re i (fJ + 2mr )

Vn EZ.

130

4. SPECTRAL DECOMPOSITION

In practice, R is found by spectrally decomposing AtA and taking its positive square root. 6 Once R is found, U can be calculated from the definition A = UR. 4.6.2. Example. Let us findthe polar decomposition of A=

(-gi

~.

We have

Theeigenvalues and eigenvectors of R2 areroutinely found to be Al = 18,

1.2 = 2,

let}

I = 2../2

(i)

~ ,

Theprojection matrices are

i../7\ 7 },

Thus,

TofindU, we note thatdetAis nonzero.Hence, A is invertible, which implies that Ris also invertible. Theinverse of Ris

The unitary mattix is simply U = AR- I

= ~ (-iI5../2 24

3i-m

3v'14\

15../2)·

It is left for thereader to verifythat Uis indeed unitary.

4.7 Real Vector Spaces The treatment so far in this chapter has focused on complex inner product spaces. The complex number system is far more complete than the real numbers. For example, in preparation for the proof of the spectral decomposition theorem, we used the existence of n roots of a polynomial of degree n over the complex field 6It is important to pay attention to the orderof the two operators: One decomposesAtA, not AAt.

4.7 REAL VECTOR SPACES

131

(this is the fundamental theorem of algebra). A polynomial over the reals, on the other hand, does not necessarily have all its roots in the real number system. It may therefore seem that vector spaces over the reals will not satisfy the nseful theorems and results developed for complex spaces. However, through a process called complexification of a real vector space, in which an imaginary part is added to such a space, it is possible to prove (see, for example, [Balm 58]) practically all the results obtained for complex vector spaces. Only the resnlts are given here. 4.7.1. Theorem. A realsymmetricoperatorhasa spectraldecompositionas stated

in Theorem4.4.6. This theorem is especially useful in applications of classical physics, which deal mostly with real vector spaces. A typical situation involves a vector that is related to another vector by a symmetric matrix. It is then convenient to find a coordinate system in which the two vectors are related in a simple mauner. This involves diagonalizing the symmetric mattix by a rotation (a real orthogonal matrix). Theorem 4.7.1 reassures us that such a diagonalization is possible. 4.7.2. Example. For a system of N point particles constituting a rigid body, the total angular momentum L = 'Lf::l mj(rj x Vi) is related to the angular frequency via L = 'Lf::l mi[rj x (w x Ii)] = 'Lf::l mj[wfj . f j

;)= (~;: ~;~ (Z t., I zx

/zy

-

fieri . W)], or

IX,) (WX) Wy.

I yz I zz

{Oz

where =

L:mjCrl- xl>. ;=1

N

N

N

i..

I yy

= L mi (r1- Y1),

I zz =

EmjCrl

- zf),

i=1

i=1

N

N

N

Ix y = -LmjXiYi,

[xz = - LmjXjZi.

i=1

i=1

I yz

= - LmiYiZi, i=1

with I x y = I yx • [xz = [zx. and I yz = flY' The 3 x 3 matrix is denoted by I and is called the moment of inertia matrix. It is symmetric,andTheorem4.7.1permits its diagonalization by an orthogonal transformation (the counterpart of a unitary transformation in a real vector space). But an orthogonal transformation in three dimensions is merely a rotation of coordinares.? Thus, Theorem 4.7.1 says that it is always possible to choose coordinate systems in which the moment of inertia matrix is diagonal. In such a coordinate system we have Lx = Ixxwx. L y = Iyywy, and L z = Izzwz, simplifying the equations considerably.

Similarly, the kineticenergyof the rigid rotatingbody, T =

N.

N

N

i=1

i=1

i=1

L !mjVr = L !miVj . (w x rj) = L !mjw. (r, x Vj) = !W. L = !wt!W,

71bis is not entirely true! There are orthogonal transformations that are composed of a rotation followed by a reflection about the origin. See Example 3.5.8.

132

4. SPECTRAL DECOMPOSITION

which in general has off-diagonal terms involving I x y and so forth, reduces to a simple IIzzw 2 III rorm: T = 2.I Ixxwx2 + 2"IIyyW 2y + '2 z·

&.

4.7.3. Example. Another application of Theorem 4.7.1 is in the study of conic sections.

The most general form of the equation of a conic section is alx

2

+ azy2 + a3xy + a4x + a5Y + a6 =

0,

where aI, ... , a6 are constants. ITthe coordinate axes coincide with the principal axes of the conic section, the xy term will be absent, and the equation of the conic section takes the familiar form, On geometrical grounds we have to be able to rotate xy-coordinates to coincide with the principal axes. We shall do this using the ideas discussed in this chapter. First, we note that the general equation for a conic section can be written in matrix form

as (x

as)

(~) + a6 = o.

The 2 x 2 matrix is symmetricand can therefore be diagonalized by means of an orthogonal matrix R. Then R/R = 1, and we can write

Let

i

a3/2) Rt = (a a2 0 Then we get

~) (;:) + (a4 a~) (;;) + a6 =

(x'

y') (aJ

"2 alx

"2 +a" " +a6 = 0 . +a2Y 4x +asy

0;

or

The cross term has disappeared. The orthogonal matrix R is simply a rotation. In fact, it rotates the original coordinate system to coincide with the principal axes of the conic section. II 4.7.4. Example. In this example we investigate conditions under which a multivariable function has a maximum or a minimum. A point a = (aI, a2, ...• an) E lRn is a maximum (minimum) of a function

f(xj, X2, ... , xn) ee f(r)

if

v flx,=a, =

(aa

f

Xl

,

aaf

x2

, ... ,

aa f

xn

= 0

) Xj=aj

4.7 REAL VECTOR SPACES

133

and for small Xi - ai, the difference fer) - /(a) is negative (positive). To relate this

difference to the topicsof this section, write the Taylorexpansionof the functionarounda keeping terms up to the second order: n al l(r)=/(a)+L(Xi-ai)(-a.) X, r=a

i=l

2 1 n ( a I ) +-2 L(Xi- ai)(Xj-aj) -a·a· XI XJ

i,j

r=a

+ ... ,

or, constructing a column vector out of 8j ea Xi - ai and a symmetric matrix Dij out of the second derivatives, we can write

1 t I(r) - I(a) = -8 08 + ... 2 because the first derivatives vanish. For a to be a minimum point of f, the RHS of the last equation must be positive for arbitrary 8. This means that D must be a positive matrix. S Thus, all its eigenvalues must be positive (Corollary 4.4.8). Similarly, we can show that for a to be a maximum point of f. -0 must be positive definite. This means that D must have negative eigenvalues. When we specialize the foregoing discussion to two dimensions, we obtainresults that are familiar from calculus. For the function f (x, y) to have a minimum, the eigenvalues of the matrix

Ixy) ( Ixx Iyx Iyy must be positive. The characteristic polynomial det (Ixx

Iyx

Ixy )

A

Iv»

=0

-A

yields two eigenvalues:

Ixx + I yy + JUxx - l yy)2 +4!A At =

2

'

Ixx + Iyy - JUxx - l yy)2 +4!ly 1.2 =

2

.

These eigenvalues will be both positive if

Ixx + I yy > jUxx:- /yy)2 +4!ly. and both negative if

Ixx + Ivv

<

-JUxx - lyy)2 +41;y.

Squaring these iuequalities and simplifying yields 2 Ixxlyy > IxY '

SNote that D is already symmetric-the real analogue of hermitian.

134

4. SPECTRAL DECOMPOSITION

which shows that fxx and /yy must have the same sign. ITthey are both positive (negative), we have a minimum (maximum).This is the familiar condition for the attainment of extrema by a function of two variables. II

Although the establishment of spectral decomposition for symmetric operators is fairly straightforward, the case of orthogonal operators (the counterpart of unitary operators in a real vector space) is more complicated. In fact, we have already seen in Example 4.5.1 that the eigenvalues of an orthogonal transformation in two dimensions are, in general, complex. This is in contrast to symmetric transformations, Think of the orthogonal operator 0 as a unitary operator.? Since the absolute value of the eigenvalues of a unitary operator is I, the only real possibilities are ± I. To find the other eigenvalues we note that as a unitary operator. 0 can be written as e A , where A is anti-hermitian (see Problem 4.22). Since hermitian conjugation and transposition coincide for real vector spaces, we conclude that A = _At, and A is antisymmetric. It is also real, because 0 is. Let us now consider the eigenvalues of A. If A is an eigenvalue of A corresponding to the eigenvector la), then (al Ala) = A (al a). Taking the complex conjugate of both sides gives (al At la) = A* (al a); but At = At = -A. because A is real and antisymmetric. We therefore have (alAla) = -A* (ala), which gives A* = -A. It follows that if we restrict A to be real. then it can only be zero; otherwise. it must be purely imaginary. Furthermore, the reader may verify that if A is an eigenvalue of A. so is -A. Therefore, the diagonal form of A looks like this: Amag

= diag(O. 0, ...• O. WI. -WI.

ifh. -ilh•. . . • Wk, -iek).

which gives 0 the following diagonal form: ' Odlag

= e Adiag = diag(e O, eO , ... " eO eifft , e-i91 , eilh , e-ifh , .•• , eiOk , e-Uh)

with el • fh•. . . . ek all real. Ilis clear that if 0 has -I as an eigenvalue, then some of the e's must equal ±1f. Separating the 1f'S from the rest of e's and putting all of the above arguments together, we get O' - diagt l , 1, ... , 1,~,e -1 -1 -1 Uft ,e-Uh ,e,e i(h -iBz, ... ,eiBm ,e-iBm) dlagN+

N_

whcre v., +N_ +2m = dimO. Getting insight from Example 4.5.1, we can argue. admittedly in a nonrigorous way. that corresponding to each pair e±iBj is a 2 x 2 matrix of the form

-Sine}) COse}

== R (e.) 2

J

(4.17)

9This can always be done by formally identifying transposition with hermitian conjugation, an identification that holds when the underlying field of numbers is real.

4.7 REAL VECTOR SPACES

135

We therefore have the following theorem (refer to [Hahn 58] for a rigorous treatment).

4.7.5. Theorem. A real orthogonal operator on a real inner product space V cannot, in general, be completely diagonalized. The closest it can get to a diagonal form is Odiag = diag(l, 1, ... ,1",-1, -1, ... , -1, R2(11j), R2(!h), ... , R2(em )) , where N+ + N_ + 2m = dim V and R2(ej) is as given in (4.17). Furthermore, the matrix that transforms an orthogonal matrix into the form above is itself an orthogonal matrix. The last statement follows from Theorem 4.5.2 and the fact that an orthogonal matrix is the real analogue of a unitary matrix.

4.7.6. Example. An interesting application of Theorem4.7.5 occursin classicalmechanics, whereit is shownthat themotionof arigidbodyconsistsof atranslation andarotation. 'The rotation is represented by a 3 x 3 orthogonal matrix. Theorem 4.7.5 states that by an appropriate choice of coordinate systems (l.e., by applying the same orthogonal transformation that diagonalizes the rotation matrix of the rigidbody), one can "diagonalize" the 3 x 3 orthogonal matrix. The "diagonal" formis

o

o

±l

or

o

cosB

sinO

-s~e) . cosB

Excluding the reflections (corresponding to -1 's) and the trivial identity rotation, we concludethat anyrotation of a rigidbody can be written as I

o (o

0 cosB

sine

-s~e) , cosO

whichis a rotation through the angle B about the (new)x-axis.

III

Combining the rotation of the example above with the translations, we obtain the following theorem.

4.7.7. Theorem. (Euler) The general motion ofa rigid body consists ofthe translation of one point of that body and a rotation about a single axis through that point.

Finally, we quote the polar decomposition for real inner product spaces.

4.7.8. Theorem. Any operator A on a real inner product space can be written as A = OR, where R is a (unique) symmetric positive operator and 0 is orthogonal.

136

4. SPECTRAL DECOMPOSITION

4.7.9. Example. Let us decompose the following matrix into its polar form:

Theprocedure is the sameas in the complexcase. Wehave

2

t

(2 -23)(23 -20) = (13 -6

R = AA= 0

witheigenvalues A.l = 1 and).2 = 16 andnormalized eigenvectors

and The projection operators are PI

= le1)(eIl = ~

G;),

-2) 1.

Thus, we have

-2) = ~ (17 -6) 1 5-6 8 . Wenote that A is invertible. Thus, Ris also invertible, and

I R- =

2~ (~ I~)'

This gives0 = AR- 1, or

It is readilyverified that 0 is indeedorthogonal.

III

Our excursion through operator algebra and matrix theory bas revealed to us the diversity of diagonalizable operators. Could it be perhaps that all operators are diagonalizable? In other words, given any operator, can we find a basis in which the matrix representing that operator is diagonal? The answer is, in general, no! (See Problem 4.27.) Discussion of this topic entails a treatment of the HamiltonCayly theorem and the Jordan canonical form of a matrix, in which the so-called generalized eigenvectors are introduced. A generalized eigenvectorbelongs to the kernel of (A - A1)m for some positive integer m. Then A is called a generalized eigenvalue. We shall not pursue this matter here. The interested reader can find such a discussion in books on linear algebra and matrix theory. We shall, however, see the application of this notion to special operators on infinite-dimensional vector spaces in Chapter 16. One result is worth mentioning at this point.

4.7.10. Proposition. If the roots of the characteristic polynomial of a matrix are all simple, then the matrix can be brought to diagonal form by a similarity (not necessarily a unitary) transformation.

4.7 REAL VECTOR SPACES

137

4.7.11. Example. As a finat example of the application of the results of this section, tet us evaluate the n-fold integral In =

100 100 dXl

-00

dX2 ...

-00

100

-'"

dXn e - L..i,j=l

-00

(4.t8)

lnijXiXj,

where the mij are elements of a real, symmetric, positive definite matrix. say M. Because it is symmetric, M can be diagonaIized by an orthogonal matrix R so that RMR t = D is a diagonal matrix whose diagonal entries are the eigenvalues, AI. ).,2, ... , An, of M, whose positive definiteness ensures that none of these eigenvalues is zero or negative. The exponent in (4.18) can be written as n

L

mijXjXj = xtMx =xtRtRMRtRx

= xttDx' = )"1X12

+ ... +).nx~2.

i,j=l

where

or, in component form,

x; =

L:j=l TijXj

for t = 1,2, ...

,n. Similarly, since x = Rtx'. it

=

follows that Xi LJ=l rjixj for i = 1,2, ... , n. The "volume element" dx! dXn is related to the primed volume element as follows:

,xn)1 dX

8 (Xl> X2 ,

dXl···dxn =

f"' 1

f

8 (xl' x 2 . · · · , x n)

I

d

I

I

I

l ···dxn es IdetJldxl ... x n'

where J is the Jacobian matrix whose ijth element is 8xi/8xj. But 8Xi

8' =rji x

[det J] = IdetR'1 = I.

j

Therefore, in terms of s', the integral In becomes

1 dx~ l°Odx~ 1 dx~e-J..IX?-J..2Xq-- ·-J..nX~2 dx~e-J..1X?) (i: (i: dx~e-J..nX~) 00

In =

00

...

-00

-00

= (/_: =

analytic definition of the determinant ofa matrix

..

-00

dx2e-J..2X22) ...

~ ~ ... ~ = "n/2

VAt V1.2

vi;.

1 = "n/2(detM)-t/2, .JA1A2"· An

because the determinant of a matrix is the product of its eigenvalues. This result can be written as

1

00

t

dnxe- X Mx = "n/2(detM)-1/2

=}

detM =

-00

which gives an analytic definition of the determinant.

nn

(J~oo d n xe- x t M X )2

,

III

138

4. SPECTRAL DECOMPOSITION

4.8 Problems 4.1. Let 11j and 11z be subspaces of V. Show that (a) dim(l1j + 11z) = dim 111 + dim 11z - dim(111 n 11z).Hint: Extend a basis of 111 n 11z to both 111 and 11z. (b) If111 + 11z = V and dim 111 + dim 11z = dim V, then V = 111 m11z. (b) If dim 111 + dim 11z > dim V, then 111 n 11z oF {OJ.

4.2. Let P be the (hermitian) projection operator onto a subspace Jye Show that 1 - P projects onto M1-. Hint: You need to show that (ml Pia) = (ml a) for arbitrary la) and 1m) E M; therefore, consider (ml P la)*, and use the hermiticity ofP. 4.3. Show that a subspace M of an inner product space V is invariant under the linear operator A if and only if M 1- is invariant under At. 4.4. Show that the intersection of two invariant subspaces of an operator is also an invariant subspace. 4.5. Let n be a permutation of the integers (I, 2, ... , n j. Find the spectrum of An, if for Ix) = (aj, az, ... ,an) E en, we define An [x) = (an(I), ... , an(n»)'

4.6. Show that (a) the coefficient of J..N in the characteristic polynomial is (_l)N, where N = dim V, and (b) the constant in the characteristic polynomial of an operator is its determinant.

translation operator

4.7. Operators A and B satisfy the commutation relation [A, B] = 1. Let Ib) be an eigenvector of B with eigenvalue J... Show that e-'fA Ib) is also an eigenvector of B, but witheigenvalue J.. + r , This is why e-'fA is called the translation operator for B. Hint: First find [B, e-'fA]. 4.8. Find the eigenvalues of an involutive operator, that is, an operator A with the property AZ = 1. 4.9. Assume that A and A' are similar matrices. Show that they have the same eigenvalues. 4.10. In each of the following cases, determine the counterclockwise rotation of the xy-axes that brings the conic section into the standard form and determine the conic section. (a) llx z + (e) 2x z (e)

3l + 6xy -

4xy -

12 = 0

3= 0

2x z + sl- 4xy - 36 = 0

(b) Sx Z - 3yz+6xy+6=O z (d) 6x + 4xy - 7 = 0

3l-

4.B PROBLEMS

139

4.11. Show that if A is invertible, then the eigenvectors of A-I are the same as those of A and the eigenvalues of A-I are the reciprocals of those of A. 4.12. Find all eigenvalues and eigenvectors of the following matrices:

Al =

A2=

A3=

(~ ~)

81 =

-2

(~ ~)

3

Cl= CI

~I)

-4 -I

I

I -I I

G ~) G D C JJ G D G D ( D I

0 I 0 I I

0

82=

0

83 =

I I I

C2 =

:

I

C3 =

:

I 0 I

4.13. Sbow that a 2 x 2 rotation matrix does not have a real eigenvalue (and, therefore, eigenvector) when the rotation angle is not an integer multiple of tt . What is the physical interpretation of this? 4.14. Three equal point masses are located at (a, a, 0), (a, 0, a), and (0, a, a). Find the moment of inertia matrix as well as its eigenvalues and the corresponding

eigenvectors. 4.15. Consider (aI, a2, ... , an) E en and define Eij as the operator that interchanges a: and a i- Find the eigenvalues of this operator. 4.16. Find the eigenvalues and eigenvectors of the operator -id/dx acting in the vector space of differentiable functions l (-00,00).

e

4.17. Show that a hermitian operator is positive if and only if its eigenvalues are positive. 4.18. What are the spectral decompositions of At, A-I, and AAt for an invertible normal operator A? 4.19. Consider the matrix

I

+ i) . 3

(a) Find the eigenvalues and the orthonormal eigenvectors of A. (b) Calcnlate the projection operators (matrices) PI andP2 and verify that Li Pi = 1 and Li AiPi = A. (c) Find the matrices.,fA, sin(nA/6), and cos(nA/6). (d) Is A invertible? If so, find the eigenvalues and eigenvectors of A-I.

140

4. SPECTRAL OECOMPOSITION

4.20. Consider the matrix

A=

(~i ~ ~}

(a) Find the eigenvalnes of A. Hint: Try A = 3 in the characteristic polynomial of

A. (b) For each A, find a basis for

M, the eigenspace associated with the eigenvalue

A. (c) Use the Gram-Schmidt process to orthonormalize the above basis vectors. (d) Calculate the projection operators (matrices) Pi for each subspace and verify that Pi 1 and AiPi A. (e) Find the matrices.,fA, sin(rrAj2), and cos(rrAj2). (f) Is A invertible? If so, find the eigenvalues and eigenvectors of A-I.

Li =

Li

=

4.21. Show that if two hermitian matrices have the same set of eigenvalues, then they are unitarily related. 4.22. Prove that corresponding to every unitary operator U acting on a finitedimensional vector space, there is a hermitian operator H such that U = exp iH. 4.23. Find the polar decomposition of the following matrices:

0)

41

2i A="fi3' (

B = ( 12i

-12i) 34

c=

'

1 0 0 1 (I i

4.24. Show that an arbitrary matrix A can be "diagonalized" as D = UAV, where U is unitary and D is a real diagonal matrix with only nonnegative eigenvalues. Hint: Consider AAt . 4.25. Show that (a) if A is an eigenvalue of an antisymmetric operator, then so is -A, and (b) antisymmetric operators (matrices) of odd dimension cannot be invertible. 4.26. Find the unitary matrices that diagonalize the following hermitian matrices: Al =

BI =

C12_i -1 0

(~I -i

-1+ -1

i) '

, , ~)

-I

-,

A2=C·

B2 =

(~.-,

;) 0 -1

,

A3 =

G -i) o '

~i) .

Warning! You may have to resort to numerical approximations for some of these.

(b

4.27. For A = 1), where x t= 0, show that it is impossible to find an invertible 2 x 2 matrix R such that RAR- 1 is diagonal. (This shows that not all operators are diagonalizable.)

4.8 PROBLEMS

141

Additional Reading 1. Alder, S. Linear Algebra Done Right, Springer-Verlag, 1996. Concise but useful discussion of real and complex spectral theory. 2. DeVito, C. Functional Analysis and Linear Operator Theory, AddisonWesley, 1990. Has a good discussion of spectral theory for finite and infinite dimensions. 3. Halmos, P. Finite Dimensional Vector Spaces, 2nd ed., Van Nostrand, 1958. Comprehensive treatment of real and complex spectral theory for operators

on finite dimensional vectorspaces.

Part II

Infinite-Dimensional Vector Spaces

_

5

_

Hilbert Spaces

The basic concepts of finite-dimensional vector spaces introduced in Chapter I can readily be generalized to infinite dimensions. The definition of a vector space and

concepts of linear combination, linear independence, basis,subspace, span, andso forth all carry over to infinite dimensions. However, one thing is crucially different in the new situation, and this difference makes the study of infinite-dimensional vector spaces bothricher and morenontrivial: In a finite-dimensional vectorspace we dealt withfinite sums; in infinite dimensions we encounter infinite sums. Thus, we have to investigate the convergence of such sums.

5.1 The Question of Convergence The intuitive notion of convergence acquired in calculus makes use of the idea of closeness. This, in tum, requires the notion of distance.' We considered such a notion in Chapter I in the context of a norm, and saw that the inner product had

anassociated norm. However, it is possible to introduce a norm on a vectorspace without aninner product. One such norm, applicable to en and JRn, was

where p is an integer. The "natural" norm, i.e., that induced on en (or JRn) by the usual inner product, corresponds to p = 2. The distance between two points 1It is possible to introduce the idea of closeness abstractly, withoutresort to the notion of distance,as is done in topology. However, distance,as appliedin vector spaces, is as abstract as we wantto get.

146

5. HILBERT SPACES

Closeness isa relative concept!

Cauchy sequence defined

depends on the particntar norm nsed. For example, consider the "point" (or vector) Ib) = (0.1,0.1, ... ,0.1) in a 1000-dimensional space (n = 1000). One can easily check that the distance of this vector from the origin varies considerably with p: IIblil = 100, IIbll2 = 3.16, libII 10 = 0.2. This variation may give the impression that there is no snch thing as "closeness", and it all depends on how one defines the norm. This is not true, becanse closeness is arelative concept: One always compares distances. A norm with large p shrinks all distances of a space, and a norm with small p stretches them. Thus, although it is impossible (and meaningless) to say that "Ia) is close to Ib)" because of the dependence of distance on p, one can always say "Ia) is closer to Ib) than [c) is to Id)," regardless of the value of p, Now that we have a way of telling whether vectors are close together or far apart, we can talk about limits and the convergence of sequences of vectors. Let us begin by recalling the definition of a Cauchy sequence

5.1.1. Definition. An infinite sequence of vectors {lai)}~l in a normed linear space V is called a Cauchy sequence if 1im;-->oo lIa; - aj II = O. j-HX)

A convergent sequence is necessarily Cauchy. This can be shown using the triangle inequality (see Problem 5.2). However, there may be Cauchy sequences in a given vector space that do not converge to any vector in that space (see the example below). Snch a convergence requires additional properties of a vector space summarized in the following deliuition. complete vector spacedefined

5.1.2. Definition. A complete vector space V is a normed linear space for which every Cauchy sequence of vectors in V has a limit vector in V. In other words, if {Iai) }~l is a Cauchy sequence, then there exists a vector la} E V such that limi-->oo lIa; - a II = o. 5.1.3. Example. 1. lll. is completewith respect to the absolute-value norm lIaII = la I. In otherwords,everyCauchysequenceof realnumbers has a limit in JR. This is provedin real analysis. 2.
II

Let {laj}}~l be a Cauchy sequence of vectors in a finite-dimensional vector space V N. Choose an orthonormal basis {lek)}f=l in VN such thatz laj} =

2Recall that one can always definean innerproduct on a finite-dimensional vectorspace. So, the existenceof orthonormal bases is guaranteed.

5.1 THE QUESTION OF CONVERGENCE

147

N W N ill Lk=l Cl k [ek) and [aj) = Lk=t Cl k [ek). Then

lIa; _ajll2 = (a; -ajla; -aj) =

11~(Clf) -Clkj») lekf

N

L (Clk;) -

=

N

Clkj))*(Cl?) - Cl?») (ek[el) =

k,l=l

L [Clf) -

j) Clk [2.

k=l

The LHS goes to zero, because the sequence is assumed Cauchy. Furthermore, all terms on the RHS are positive. Thus, they too must go to zero as i, j --> 00. By n) the completeness of C, there must exist Clk E C such that limn .... oo Clk = Clk for k = 1,2, ... , N. Now consider [a) E VN given by la) = Lf=l Clk lek). We claim that [a) is the limit of the above sequence of vectors iu VN. Indeed,

We have proved the followiug: allfinite-dimensional vector spaces are complete

5.1.4. Proposition. Every Cauchy sequence in a finite-dimensional inner product space overC (or R) is convergent. In otherwords, everyfinite-dimensional complex (or real) inner product space is complete with respect to the norm induced by its inner product.

Thenext example showshow important the word "finite" is. 5.1.5. Example. Consider {fk}~l' the infinitesequence of continuous functions defined

iu the interval [-I, +1] by

Ik(x) =

{

Ilk::ox::ol, -llk::o x::o Ilk, -I::ox::o -Ilk.

if

I

(kx

+ 1)/2

o

if if

This sequence belongsto eO(_l, 1), the inner product space of continuous functions with its usual inner product: (II g) = J~l !*(x)g(x) dx. It is straightforwardto verify that

IIIk - /;11 2

= J~ll!k(x) -

/;(x)1 2dx

.

,0. Therefore, the sequence is Cauchy.

k,j"""'*oo

However. thelimit of this sequence is (see Figure 5.1)

I(x) =

{I

o

if

0< x < I,

if

-1
which is discontinuous at x = 0 andtherefore does not belong to the space in which the original sequence lies. III

148

5. HI LBERT SPACES

y

o Figure 5.1 Thelimit of the sequence of the continuous functions function that is 1 for x > 0 and0 for x < O.

Banach space

Hilbert space defined

Ik is a discontinuous

We see that infinite-dimensional vector spaces are not generally complete. It is a nontrivial taskto show whether or not a given infinite-dimensional vector space is complete. Any vector space (finite- or infinite-dimensional) contains all finite linear combinations of the form I:7~1 a; lai) when it contains all the lai)'s. This follows from the very definition of a vector space. However, the sitnation is different when n goes to infinity. For the vector space to contain the infinite sum, firstly, the meaning of such a sum has to be clarified, i.e., a norm and an associated convergence criterion needs to be put in place. Secondly, the vector space has to be complete with respect tn that norm. A complete normed vector space is called a Banach space. We shall not deal with a general Banach space, but only with those spaces whose norms arise natnrally from an inner product. This leads to the following definition: 5.1.6. Definition. A complete inner product space, commonly denoted by 11:, is called a Hilbert space. Thus, all finite-dimensional real or complex vector spaces are Hilbert spaces. However, when we speak of a Hilbert space, we shall usually assume that it is infinite-dimensional. It is convenient to use orthonormal vectors in stndying Hilbert spaces. So, let us consider an infinite sequence (lei) }~1 of orthonormal vectors all belonging to a Hilbert space 11:. Next, take any vector If) E 11:, construct the complex numbers fi = (ei If), and form the sequence of vectors'' n

Ifn) =

L fi lei)

for n = 1,2, ...

(5.1)

;=1

3WecanconsiderIfn) asan"approximation" to If), becausebothshare thesamecomponents alongthesamesetof orthonormal vectors. The sequence of orthonormal vectorsactsverymuchas a basis.However, to be a basis, an extracondition mustbe met. We shalldiscussthis condition shortly.

5.1 THE QUESTION OF CONVERGENCE

For the pair of vectors

149

If) and IJ.,}, the Schwarz inequality gives (5.2)

where Equation (5.1) has been used to evaluate Unl fn}. On the other hand, taking the inner product of (5.1) with (fl yields

n n n Ulfn) = Lfdflei) = Lfd,' = Ll.fil2. ;=1

i=1

Parseval inequality

;=1

Substitution of this in Equation (5.2) yields the Parseval inequality: n

L I.ti 1 :s UI f} . 2

(5.3)

;=1

This conclusion is true for arbitrarily large n and can be stated as follows:

Bessel inequality

5.1.7. Proposition. Let Ilei)}?::t be an infinite set of orthonormal vectors in a Hilbert space, 1[, Let If) E 11: and define complex numbers Ii = (eil f). Then the Besselinequality holds: E?::l Ifil 2 s UI f). The Bessel inequality shows that the vector n

00

"L- Ii

lei)

i=1

== n--+ooLlim "Ii

lei)

i=l

converges; that is, it has a finite norm. However, the inequality does not say whether the vector converges to If). To make such a statement we need completeness: complete orthonormal sequence ofvectors

5.1.8. Definition. A sequence oforthonormal vectors I lei)}?::! in a Hilbert space 11: is called complete if the only vector in 11: that is orthogonal to all the lei} is the zero vector. This completeness property is the extra condition alluded to (in the footuote) above, and is what is required to make a basis. 5.1.9. Proposition. Let Ilei}}?::! be an orthonormal sequence in 11:. Then the following statements are equivalent:

1. Ilei}}?::! is complete.

2. If) = E~! [ei} (eil f)

v If)

E 1[,

3. E~l [ei} (eil = 1.

4. UI g)

=

E?::l UI ei) (eil g)

v If), Ig}

E

11:.

150

5. HILBERT SPACES

v

If) E 1f.

Proof We shall prove the implications 1=}2 =} 3 =} 4 =} 5 =} I. I =} 2: It is sufficient to show that the vector 11ft) == If) - I:f;:! lei) (eil f) is otthogonal to all the Iej):

(ejl1ft)

= (ejl f)

8ij 00 , . . - ' - - ,

-

L (ejl ei) (eil f) = O. ;=1

2 =} 3: Since If) = llf) = I:f;:j(lei) (eil) If) is true for all If) E 1C, we must have 1 = I:f;:I!ei)(eil. 3 =} 4: (fl g) = (flllg) = (fl (I:f;:I!ei) (eil) Ig) = I:f;:j (fl ei)(eil g). 4 =} 5: Let Ig} = If) in statement 4 and recall that (fl ei) = (eil f)*. 5 =} 1: Let If) be otthogonal to all the lei). Then all the terms in the sum are zero implying that 1If11 2 = 0, which in tum gives If) = 0, because only the zero vectorhas a zero norm. 0 Parseval equality; generalized Fourier

The equality

coefficients

00

IIff

= (fl f) = L

00

I (eil f) 12 =

;=1

completeness relation

L

2

Ifil ,

fi = (eil f) ,

(5.4)

i=1

is called the Parseval equality, and the complex numbers Fourier coefficients. The relation

Ji are called generalized

00

1=

L

lei} (eil

(5.5)

;=1

is called the completeness relation. basis for Hilbert spaces

5.1.10. Definition. A complete orthonormal sequence (lei) Jf;:! in a Hilbert space 1Cis called a basis of1C.

5.2 The Space of Square-Integrable Functions Chapter 1 showed that the collection of all continuous fuoctions defined on an interval [a, b] forms a linear vector space. Example 5.1.5 showed that this space is not complete. Can we enlarge this space to make it complete? Since we are interested in an inner product as well, and since a natural inner product for functions is defined in terms of integrals, we want to make sure that our fuoctions are integrable. However, integrability does not require continuity, it only requires piecewise continuity. In this section we shall discuss conditions under which the

5.2 THE SPACE OF SQUARE-INTEGRABLE FUNCTIONS

151

space of functions becomes complete. An important class of functions has already been mentioned in Chapter I. These functions satisfy the inner product given by

l

(gl f) = square-integrable functions

If g(x) =

f

b

g*(x)!(x)w(x) dx.

ix), we obtain

(fl f) =

l

b

1!(x)1

2w(x)

dx.

(5.6)

Functions for which such an integral is defined are said to be square-integrable.

David Hilbert (1862-1943), the greatest mathematician of this century, receivedhis Ph.D. from the University of Konigsberg and was a member of the staff there from 1886 to 1895.ln 1895 he was appointed to the chair of mathematics at the University of Gottingen, wherehe continued to teachfor therestof his life. Hilbert is one of that rare breedof late 19th-century mathematicians whose spectrum of expertise covered a wide range, withformal settheory atoneendandmathematical physicsatthe other.He did superb work in geometry,algebraic geometry,algebraic number theory, integral equations, and operator theory. The seminaltwo-volume book Methoden der mathematische Physik by R. Courant, still one of thebestbookson the subject, was greatly influenced by Hilbert. Hilbert's work in geometry hadthegreatest influence in that area sinceEuclid. A systematicstudy of theaxioms of Euclidean geometry led Hilbert to propose 21 suchaxioms, and he analyzed their significance. He published Grundlagen der Geometrie in 1899, puttiug geometry on a formal axiomatic foundation. His famous 23 Paris problems challenged (and stilltodaychallenge) mathematicians to solve fundamental questions. It was late in his career that Hilbert turned to thesubject for whichhe is mostfamous among physicists. A lecture by Erik Holmgren in 1901 on Fredholm's work on integral equations, whichhad already beenpublished in Sweden, aroused Hilbert's interest in the subject. DavidHilbert, having established himselfas theleading mathematician of his time by his work on algebraic numbers, algebraic invariants, andthefoundations of geometry, nowturned his attention to integral equations.He saysthat aninvestigation of thesubject showed him that it wasimportant forthetheory of definite integrals, forthedevelopment of arbitrary functions in series (of specialfunctions ortrigonometric functions). forthetheory of linear differential equations, for potential theory. andfor thecalculus of variations. He wrote a series of six papers from 1904to 1910andreproduced them in hisbook Grundzuge einer allgemeinenTheone der linearenIntegralgleichungen (1912). During thelatter part of thisworkhe applied integral equations to problems of mathematical physics. Itis saidthat Hilbert discovered thecorrect fieldequation forgeneral relativity in 1915 (one year before Einstein) usingthe variational principle, butnever claimed priority. Hilbert claimed that he worked best out-of-doors. He accordingly attached an 18-foot blackboard to his neighbor's wall andbuilta covered walkway there so that he couldwork outside in any weather. He wouldintermittently interrupt his pacing andhis blackboard

152

5. HILBERT SPACES

computations with a few turns around the rest of the yard on his bicycle, or he wouldpull some weeds, or do some garden trimming. Once, when a visitorcalled, the maidsenthim to the backyard and advised thatif the master wasn't readily visible at the blackboard to look for him up in one of the trees. Highlygiftedandhighlyversatile, DavidHilbert radiated overmathematics a catching optimism anda stimulating vitality that canonly be called"thespirit of Hilbert," Engraved on a stone marker set over Hilbert's grave in Gottingen are the master's own optimistic words: "WiT mussenwissen. WiT werden wissen." ("We must know. We shallknow.")

The space of square-integrable functions over the interval [a, b] is denoted by .c~(a, b). In this notation L stands for Lebesgue, who generalized the notion of the ordinary Riemann integral to cases for which the integrand could be highly discontinuous; 2 stands for the power of j (x) in the integral; a and b denote the limits of integration; and w refers to the weight function (a strictly positive realvalued function). When w(x) = I, we use the notation .c 2(a , b). The significance of .c~(a, b) lies in the following theorem (for a proof, see [Reed 80, Chapter III]): £'~(a, b) is complete

5.2.1. Theorem. (Riesz-Fischer theorem) The space .c~(a, b) is complete.

A complete infinite-dimensional ioner product space was earlier defined to be a Hilbert space. The following theorem shows that the number of Hilbert spaces is severely restricted. (For a proof, see [Frie 82, p. 216].) all Hilbert spacesare alike

5.2.2. Theorem. All infinite-dimensional complete inner product spaces are isomorphic to .c~(a, b). .c~ (a, b) is defined in terms offunctions that satisfy Equation (5.6). Yet anioner

J:

g*(x)j(x)w(x) dx: Are such integrals product involves integrals of the form well-defined and finite? Using the Schwarz inequality, which holds for any ioner product space, finite or infinite, one can show that the integral is defined. The isomorphism of Theorem 5.2.2 makes the Hilbert space more tangible, because it identifies the space with a space of functions, objects that are more familiar than abstract vectors. Nonetheless, a faceless function is very little improvement over an abstract vector. What is desirable is a set of concrete functions with which we can calculate. The following theorem provides such functions (for a proof, see [Sinon 83, pp. 154-161]). 5.2.3. Theorem. (Stone-Weierstrass approximation theorem) The sequence oj functions (monomials) {x k }, where k = 0, 1,2, ... .forms a basis oj .c~(a, b). Thus, any function j can be written as j(x) = L:~oakxk. Note that the {x k } are not orthonormal but are linearly independent. If we wish to obtain an orthononnal-or simply orthogonal-linear combination of these vectors, we can use the Gram-Schmidt process. The result will be certain polynomials, denoted by Cn(x), that are orthogonal to one another and span .c~(a, b).

5.2 THE SPACE OF SQUARE-INTEGRABLE FUNCTIONS

153

Such orthogonal polynomials satisfy very useful recurrence relations, which we now derive. In the following discussion P,k(X) denotes a generic polynomial of degree less than or equal to k. For example, 3x s - 4x2 + 5, 2x + I, -2.4x4 + 3x 3 - x 2 + 6, and 2 are all denoted by p,s(x) or p,s(x) or p,S9(X) because they all have degrees less than or equal to 5, 8, and 59. Since a polynomial of degree less than n can be written as a liuear combination of Ck(X) with k < n, we have the obvious property b Cn(x)P,n_l(x)w(x)dx = O.

1

(5.7)

Let k m and k~ denote, respectively, the coefficients of x m andx m- 1 in Cm(x), and let h m = 1b[Cm(x)fw(X)dX.

(5.8)

The polynomial CnH (x) - (knHI kn)xCn(x) has degree less than or equal to n, and therefore can be expanded as a linear combination of the Cj (x): (5.9) Take the inner product of both sides of this equation with Cm(x): b 1a

~+11b Cn+l(x)Cm(x)w(x)dx - xCn(x)Cm(x)w(x)dx kn

n

= I:aj j=O

a

(b

I,

Cj(x)Cm(x)w(x) dx.

a

The first integral on the LHS vanishes as long as m :0 n; the second integral vanishes if m :0 n - 2 [if m :0 n - 2, then xCm(x) is a polynomial of degree n - 1]. Thus, we have n

I:aj

r

I,

j=O

Cj(x)Cm(x)w(x)dx = 0

form :On-2.

a

The integral in the sum is zero unless j = m, by orthogonality. Therefore, the sum reduces to

b am 1

[Cm(x)]2 W(x) dx = 0

form:On-2.

Sincetheintegral is nonzero, we conclude that am = 0 for m = 0, 1,2, ... , n - 2, and Equation (5.9) reduces to kn+! Cn+l(X) - --xCn(x) = an-lCn-!(X) +anCn(x).

kn

(5.10)

154

5. HI LBERT SPACES

It can be shown that if we define an

s;

k n+l = --, k

(5.11)

hn-l an-l

n

a recurrence relation lor orthogonal polynomials

an

Yn=-----,

then Equation (5.10) can be expressed as

+ f3n)Cn(x) + VnCn-l(X),

C n+l(X) = (anx

(5.12)

or (5.13) Other recurrence relations, involving higher powers of x, can be obtained from the one above. For example, a recurrence relation involving x 2 can be obtained by multiplying both sides of Equation (5.13) by x and expanding each term of the RHS using that same equation. The result will be

(5.14)

5.2.4. Example. As anapplication of the recurrence relations above, let us evaluate t,

==

l

b

xCm(x)Cn(x)w(x)dx.

Substituting (5.13) in the integralgives lj

= -1

an

l

b

a

~

- --" IXn

Cm(x)Cn+l(x)w(x)dx - --" P

l

an

l

b

Cm(x)Cn(x)w(x)dx

a

b

Cm(X)Cn_l (x)w(x) dx.

a

Wenow use the orthogonality relations among the Ck(X) to obtain

1 lj = ~8m,n+l an

l

a

Yn m,n-l - -8 IXn

b

. P C~(x)w(x)dx---"8mn

l

an

b

l

a

b

C~(x)w(x)dx

2 Cm(x)w(x)dx

a

1- 8m,n+l =(Clm-l

Pm Ym+l 8mn - - 8m,n- l ) hm•

am

Cl m+l

5.2 THE SPACE OF SQUARE-INTEGRABLE FUNCTIONS

155

or if m = n

hml Ol m_ 1 I 1-

+ I,

= n, if m = n -

j

if m

-{3mhm/am -0Ym+l hm/am+l

III

1,

otherwise.

5.2.5. Example. Let us find the orthogonalpolynomialsforming a basis of £.2(-I, + I), which we denoteby Pk(X), where k is the degreeof the polynomial.Let Po(x) = L Tofind PI (x), write PI (x) = ax +b, and determinea andb in such a way that PI (x) is orthogonal to Po(x):

0=1

-I

1

1

1 PI (x) Po(x) dx =

-I

(ax

+ b) dx =

!ax21~1 + 2b = 2b.

So one of thecoefficients, b, is zero. To find the other one; we need some standardization procedure.We "standardize" Pk(X) by requiring that Pk(l) = I Vk: For k = I this yields a x 1 = I,ora = 1. so that PI (x) =X. We can calculate P2(X) similarly:Write P2(X) = ax 2 + bx + c, impose the condition that it be orthogonalto both PI (x) and Po(x), and enforce the standardizationprocedure. All this will yield 0=

1

1 P2(x) Po(x) dx = -a 2 -I 3

+ 2c,

1

1

0=

-I

2 P2(x) PI (x) dx = -b,

3

and P2(1) = a+b+c = L Thesethree equationshavethe unique solution a = 3/2, b = 0, c = -1/2. Thus, P2(X) = !(3x 2 - I). These are the first three Legendre polynomials, which arepart of a larger group of polynomials to be discussed in Chapter 7. II

5.2.1

Orthogonal Polynomials and Least Squares

The method of least squares is no doubt familiar to the reader. In the simplest procedure, one tries to find a linear function that most closely fits a set of data. By defioition, "most closely" means that the sum of the squares of the differences between the data points and the corresponding values of the linear function is minimum. More generally, one seeks the best polynomial fit to the data. We shall consider a related topic, namely least-square fitting of a givenfunction with polynomials. Suppose f(x) is a function defined on (a, b). We wantto find a polynomial that most closely approximates f. Write such a polynomial as p(x) = L:Z=o akx k, where the ak's are to be determined such that S(ao, 011, ... , an) ea

l

b [f(x) - 010 - alx - ... - anx

n]2

dx

is a minimum. Differentiating S with respect to the ak's and setting the result equal

to zerogives

as = 0=-.

oaJ

l

a

b

n ] dx, 2(-x i ) [ f(x) - I>kXk k=O

158

5. HILBERT SPACES

(a) II! ± gil = II!II + IIgll· (b) II! + gll2 + II! - gll2 = 2(1If11 + IIgll)2. (c) Using parts (a), (b), and Theorem 1.2.8, show that .c1(R) is not an inner product space. This shows that not all norms arise from an inner product.

5.6. Use Equation (5.10) to derive Equation (5.12). Hint: To find an, equate the coefficients of x" on both sides of Equation (5.10). To find an-I, multiply both sides of Equation (5.10) by Cn_IW(X) and integrate, using the definitions of k n, k~, andhno

5.7. Evaluate the integral

f:

x 2Cm(x)Cn (x)w(x) dx.

Additional Reading 1. Boccara, N. Functional Analysis, Academic Press, 1990. An application oriented book with many abstract topics related to Hilbert spaces (e.g., Lebesgue measure) explained for a physics audience. 2. DeVito, C. Functional Analysis and Linear Operator Theory, AddisonWesley, 1990. 3. Reed, M., and Simon, B. FunctionalAnalysis, Academic Press, 1980. Coauthored by a mathematical physicist (B.S.), this first volume of a four-volume encyclopedic treatise on functional analysis and Hilbert spaces bas many examples and problems to help the reader comprehend the rather abstract presentation. 4. Zeidler, E. Applied Functional Analysis, Springer-Verlag, 1995. Another application-oriented book on Hilbert spaces suitable for a physics audience.

6,

_

Generalized Functions

Once we allow the number of dimensions to be infinite, we open the door for numerous possibilities that are not present in the finite case. One such possibility arises because ofthe variety ofinfinities. We have encountered two types of infinity in Chapter 0, the countable infinity and the uncountable infinity. The paradigm of the former is the "number" of integers, and that of the latter is the "number" of real numbers. The nature of dimensionality of the vector space is reflected in the components of a general vector, which has a finite number of components in a finite-dimensional vector space, a countably infinite number of components in an infinite-dimensional vector space with a countable basis, and an uncountably infinite number of components in an infinite-dimensional vector space with no countable basis.

6.1 Continuous Index To gain an understanding of the nature of, and differences between, the three types of vector spaces mentioned above, it is convenient to think of components as functions of a "counting set." Thus, the components Ji of a vector If) in an N-dimensional vector space can be thought of as values of a function f defined on the finite set {I, 2, ... , N}, and to emphasize such functional dependence, we write f (I) instead of fi- Similarly, the components Ji of a vector If) in a Hilbert space with the countable basis B = (lei)}~l can be thought of as values of a function f : N --> C, where N is the (infinite) set of natural numbers. The next step is to allow the counting set to be uncountable, i.e., a continuum such as the real numbers or an interval thereof. This leads to a "component" of the form f (x) corresponding to a function f : lR --> C. What about the vectors themselves? What sort of a basis gives rise to such components?

160

6. GENERALIZED FUNCTIONS

Because of the isomorphism of Theorem 5.2.2, we shall concentrate on .c~ (a, b). In keeping with our earlier notation, let {lex) lxeJR denote the elements of an orthonormal set and interpret f(x) as (ex I f). The inner product of .c~(a, b)

can now be written as (glf)

=

l

b

g*(x)f(x)w(x)dx

=

l

= (gl ( [ lex) w(x) (ex I dX)

b

(glex)(exl f) w(x)dx

i».

The last line suggests writing

l

b

lex) w(x) (ex I dx = 1.

In the physics literature the "e" is ignored, and one writes Ix) for lex). Hence, we completeness relation fora continuous index

obtain the completeness relation for a continuous index:

l

l

b

or

[x) w(x) (xl dx = 1,

b

[x) (xl dx = 1,

(6.1)

where in the secoud integral, w(x) is set equal to unity. We also have If)

=

(l

b

Ix) w(x) (xl dx) If)

=

l

b

f(x)w(x) [x) dx,

(6.2)

which shows how to expand a vector If) in terms of the Ix)'s. Take the inner product of (6.2) with (x']:

l

(xii f) = f(x ') =

b

f(x)w(x) (xii x) dx

where x' is assumed to lie in the interval (a, b), otherwise f(x ') = 0 by definition. This equation, which holds for arbitrary f, tells us inuuediately that w(x) (xii x) is no ordinary function of x and x'. For instance, suppose f(x ') = O. Then, the result of integration is always zero, regardless of the behavior of f at other points. Clearly, there is an infinitude of functions that vanish at x', yet all of them give the same integral! Pursuing this line of argument more quantitatively, one can show that w(x) (x'] x) = Oifx oF x', w(x) (xl x) = 00, w(x) (x'] x) is anevenfunctiou of x - x', and w (x) (x'] x) dx = 1. The proof is left as a problem. The reader may recognize this as the Dirac delta function

J:

Dirac delta function

8(x - x') = w(x) (x'] x) , which, for a function

l

a

b

f

(6.3)

defined on the interval (a, b), has the following property.' I

f(x)8(x _ x') dx = {f(x

0

)

if x E (a, b), ifxrt(a,b).

1For an elementary discussionof the Diracdeltafunction with manyexamplesof its application, see [Hass 99].

(6.4)

6.1 CONTINUOUS INDEX

! _._._._._._._._.__

!,

._~_._._---_._._._._._._.+-

............ -

········__·"·__····j···_--_·_--------·----i···_·, "

,. ,

J .... ~!

e-,..........

-0.5

."

_._~_

\

....... _.,..

j

i

:

o

0.5

I

.•.......... -

.

..__ ._._-_._._._.~._---_._._._._._._._._~ .._. ........ '

i '/; ----f--- _._._._.- ······1-

_.J .. -.- , ,

161

.,

.......

.j

,, ....... , .. ............... ,,;

---------.""-j

~~.-

!

1.5

........ -

i 2

2.5

Figure 6.1 The Gaussian bell-shaped curve approaches the Dirac delta function as the width of the curve approaches zero. The value of € is 1 for the dashed curve, 0.25 for the heavy curve and 0.05 for the light curve.

Written in the form (xii x) = 8(x -x')fw(x), Equation (6.3) is the generalization of the orthonormality relation of vectors to the case of a continuous index. The Dirac delta function is anything but a "function." Nevertheless, there is a well-developed branch of mathematics, called generalized function theory or functional analysis, studying it and many other functions like it in a highly rigorous fashion. We shall only briefly explore this territory of mathematics in the next section. At this point we simply mention the fact that the Dirac delta function can be represented as the limit of certain sequences of ordinary functions. The following three examples illustrate some of these representations.

6.1.1. Example. Consider a Gaussian curve whose width approaches zero at the same time that its height approaches infinity in such a way that its area remains constant. In the infinite limit, we obtain the Dirac delta function. In fact, we have 8 (x -'- x') = limE-+o

~e-(X-XI)2/E. In the limit of E ---+ 0, the height of this Gaussian goes to infin-

vH'

ity while its width goes to zero (see Figure 6.1). Furthermore, for any nonzero value of E, we can easily verify that

This relation is independent of E and therefore still holds in the limit E -+ O. The limit of the Gaussian behaves like the Dirac delta function. II

6.1.2. Example. Consider the function DT (x - x') defined as DT(X -x')

== _I

jT e' x-x

2" -T

O(

')fdt.

162

6. GENERALIZED FUNCTIONS

o

-5

5

Figure 6.2 The function sin T xfx also approaches the Dirac delta function as the width of the curve approaches zero. The value of T is 0.5 for the dashed curve. 2 for the heavy curve, and 15 for the light curve.

The integral is easily evaluated, with the result

DT(X - x') .

1 e «x-x')t i=l sinT(x -x') .

= -2rr

T

l

i(x -x')

-T

7r

x -x'

The graph of DT (x - 0) as a function of x for various values of T is shown in Figure 6.2. Note that the width of the curve decreases as T increases. The area under the curve can be calculated:

1

00

I

,

DT(X-x)dx=-

-00

7C

1

00

-00

sin T(x - x') 1 , dx=X -x tt

1

00

sin y --dy=l.

-00 Y '---'

="

Figure 6.2 shows that DT (x - x') becomes more and more like the Dirac delta function as T gets larger and larger. In fact, we have

~(x _ x')

=

lim

I sin T(x - x').

T--+oo JC

(6.5)

X - x'

To see this, we note that for any finite T we can write

DT(X _ x') = T sin T(x - x'). x T(x - x') Furthermore, for values of x that are very close to x',

T(x -x') -+ 0

and

sin T(x - x')

T(x-x')

-+ .

I

.

Thus, for such values of x and x', we have Dr(x - x') ~ (T /Jr), which is large when T is large. This is as expected of a delta function: ~ (0) = 00. On the other hand, the width

6.1 CONTINUOUS INDEX

163

of Dr (x - x') around x' is given, roughly, by the distance between the points at which DT (x - x') drops to zero: T (x - x') = ±1l'. or X - x' = ±1l'/ T. This width is roughly Ax = 2n / T. whichgoesto zero as T grows.Again,this is as expectedof the deltafunction. III

The preceding example suggests auother representation of the Dirac delta function:

joo.( ') e' x-x 'dt.

I 8(x - x') = 2". step function or B

(6.6)

-00

6.1.3. Example. A third representation of the Dirac delta function involves the step function e(x - x'), which is defined as

function I

B(x -x)

==

{O

if x e x', 1 if x>x'

and is discontinuous at x = x', We can approximate this step function by many continuous functions, such as TE (x - x') defined by

T,(x-x')== [

0

if X~X'_,.

2~(X-x'+')

if x'-,

1

if x ~ x' +€,

~x~x'+,.

where e is a small positive number as shown in Figure 6.3. It is clear that

B(x-x')= tiro T,(x-x'). ,--+0

Now let us consider the derivative of Te (x - x') with respect to x:

at;

,

-(x-x)= dx

r 1 2,

0

if x
€ < X <

x' + E,

if x > x' + E.

We note that the derivative is not defined at x = x' - e and x = x' + E. and that dT€/ dx is zero everywhere except when x lies in the interval (x' - E, x' + f), where it is equal to 1/(2f) and goes to infinity as e .....,. O.Here again we see signs of the delta function. In fact, we also note that

1

00

-00

t:

l

X

'+E -dx (dT,) (dT,) 1 dx = dx = = I. dx x' -, dx x'_, 2<

It is not surprising, then, to find that lim€......,)oo dT€ (x - x') = 8(x - x'). Assuming that the

8function as derivative 01 B function

dx

interchange of the order of differentiation and the limiting process is justified, we obtain

the importantidentity d

I

I

-B(x - x) = 8(x - x). dx

(6.7)

III

164

6. GENERALIZED FUNCTIONS

y 1

o Figure 6.3

x X'-E

x' x'+E

The step function, or 8-function, shown in the figure has the Dirac delta

function as its derivative.

Now that we havesomeunderstanding of one continuous index,we cangeneralizetheresults to several continuous indices. In theearlier discussion we lookedat j (x) as thexthcomponent of some abstract vector If). For functions of n variables, we can think of j(X!, ... , xn) as the component of an abstract vector If) along a basis vector Ix!, ... , x n }.2 This basis is a direct generalization of one continuous index to n. Then j(XI, ... , x n) is defined as j(X!, ... , x n) = (Xl, ... , Xn If). If the region of integration is denoted by n, and we use the abbreviations

thenwe canwrite

In = In

If) = j(r')

dnxj(r)w(r) [r},

In

dnxj(r)w(r) (r'] r),

(r'l r) w(r)

dnx [r) w(r) (r] = 1,

= 8(r -

r'),

(6.8)

where d" X is the "volume" element and n is the region of integration of interest. For instance, if the region of definition of the functions under consideration is the surface of the unit sphere, then [with w(r) = 1], one gets

r

dq, fan sinO so 10, q,HO, q,1

=1.

(6.9)

2Do not confusethiswith an u-dimensional vector. In fact, the dimension is a-fold infinite: each xi countsone infinite set of numbers!

6.2 GENERALIZED FUNCTIONS

165

This will be used in our discussion of spherical harmonics in Chapter 12. An important identity using the three-dimensional Dirac delta function comes from potential theory. This is (see [Bass 99] for a discussion of this equation)

, V 2 ( -1- ) = -411"8(r-r).

(6.10)

Ir-r'l

6.2 Generalized Functions Paul Adrian Maurice Dirac discovered the delta function in the late 1920s while investigating scaltering problems in quantum mechanics. This "function" seemed to violate most properties of other functions known to mathematicians at the time. Furthermore, the derivative of the delta function, 8' (x - x') is such that for any ordinary function f(x),

i:

f(x)8'(x - x')dx

=-

i:

f'(x)8(x - x') dx

= - f'(x').

We can define 8' (x - x') by this relation. In addition, we can define the derivative of any function, including discontinuous functions, at any point (including points of discontinuity, where the usual definition ofderivative fails) by this relation. That is, if cp(x) is a "bad" function whose derivative is not defined at some point(s), and f(x) is a "good" function, we can define the derivative of cp(x) by

i:

f(x)cp'(x)dx

== -

i:

f'(x)cp(x)dx.

The integral on the RHS is well-defined. Functions such as the Dirac delta function and its derivatives of all orders are not functions in the traditional sense. What is common among all ofthem is that in most applications they appear inside an integral, and we saw in Chapter 1 that integration can be considered as a linear functional on the space of continuous functions. 11 is therefore natura1to describe such functions in terms of linear functionals. This idea was picked up by Laurent Schwartz in the 1950s who developed it into a new branch of mathematics called generalized functions, or distributions. A distribution is a mathematical entity that appears inside an integral in conjunction with a well-behaved test function-which we assume to depend on n variables-such that the result of integration is a well-defined number. Depending on the type of test function used, different kinds of distributions can be defined. If we want to include the Dirac delta function and its derivatives of all orders, then the test functions must be infinitely differentiable, that is, they must be oo functions on R n (or C"), Moreover, in order for the theory of distributions to be mathematically feasible, all the test functions must vanish outside a finite "volume" of R n (or Cn ).3 One common notation for such functions is e1i'(Rn) or e1i'(cn)

e

3Such functions are saidto be of compact support.

166

6. GENERALIZED FUNCTIDNS

(F stands for "finite"). The definitive property of distributions concerns the way they combine with test functions to give a number. The test functions used clearly form a vector space over R or C. In this vector-space language, distributions are liuear functionals. The linearity is a simple consequence of the properties of the integral. We therefore have the following definition of a distribution.

generalized functions and distributions defined

6.2.1. Definition. A distribution, or generalizedfunction, is a continuous" linear functional on the space ef"(Rn) or ef"(Cn). Iff Eef" and


6.2.2. Example. An ordinary (continnons)functinn g coo be thonght of as a specialcase of a distribution, The linear functional g : ef"(R) --+ Ris simpty defined by (g, f) ea g[f] =

1:"'00 g(x)f(x) dx.

III

6.2.3. Example. An interesting application of distributions (generalized functions) occurs whenthenotionof density is generalized to include notonly(smooth) volumedensities, butalso point-like,linear, andsurface densities. Apointcbargeq locatedat rO coobethougbtofas having acharge densityper) = q8(rrO)'In the language oflinear functionals,we interpret p as a distribution, p : ef" (E 3) --+ E, whichfor anarbitrary function f gives p[f]

= (p,

f)

= qf(ro)·

(6.11)

The deltafunction character of p can be detected fromthis equation by recalling that the LHSis

On the RHS of this equation, the only volume element that contributes is the one that contains the point to: all the rest contribute zero. As b.Vi -+ 0, the only way that the RHS can give a nonzero number is for p(rO)!(rO) to be infinite. Since! is a well-behaved

function, p(rO) must be infinite, implying that per) acts as a delta function. This shows that the definitionof Equation (6.11) ieads to a delta-functionbehavior for p. Similarlyfor linearandsurface densities. III 4See [Zeidler, 95], pp. 27. 156-160, for a formal definition of the continuity of linear functionals.

6.2 GENERALIZED FUNCTIONS

167

The example above and Problems 6.5 and 6.6 suggest that a distribution that confines an integral to a lower-dimensional space mnst have a delta function in its definition.

"The amount of theoretical ground one has to cover before being able to solve problems of real practical value is rather large, but this circumstance isan inevitable consequence ofthe fundamental part played by transformation theory and is likely to become more pronounced inthe theoretical physics of thefuture." P.A.M. Dirac (1930)

"Physical Laws should have mathematical beauty." This statement was Dirac's response to the question of his philosophy of physics, posed to him in Moscow in 1955. He wrote it on a blackboard that is still preserved today. Paul Adrien Maurice Dirac (1902-1984), was born in 1902 in Bristol, England, of a Swiss, French-speaking father and an English mother. His father, a taciturn man who refused to receive friends at home, enforced young Paul's silence by requiring that only French be spoken at the dinner table. Perhaps this explains

Dirac's later disinclination toward collaboration and his general tendency to be a loner in most aspects of his life. The fundamental nature of his work made the involvement of students difficult, so perhaps Dirac's personality was well-suited to his extraordinary accomplishments. Dirac went to Merchant Venturer's School, the public school where his father taught French, and while there displayed great mathematical abilities. Upon graduation, he followed in his older brother's footsteps and went to Bristol University to study electrical engineering. He was 19 when he graduated Bristol University in 1921. Unable to find a suitable engineering position due to the economic recession that gripped post-World War I England, Dirac accepted a fellowship to study mathematics at Bristol University. This fellowship, together with a grant from the Department of Scientific and Industrial Research, made it possible for Dirac to go to Cambridge as a research student in 1923. At Cambridge Dirac was exposed to the experimental activities of the Cavendish Laboratory, and he became a member of the intellectual circle over which Rutherford and Fowler presided. He took his Ph.D. in 1926 and was elected in 1927 as a fellow. His appointment as university lecturer came in 1929. He assumed the Lucasian professorship following Joseph Larmor in 1932 and retired from it in 1969. Two years later he accepted a position at Florida State University where he lived out his remaining years. The FSU library now carries his name. In the late 1920s the relentless march of ideas and discoveries had carried physics to a generally accepted relativistic theory of the electron. Dirac, however, was dissatisfied with the prevailing ideas and, somewhat in isolation, sought for a better formulation. By 1928 he succeeded in finding an equation, the Dirac equation, that accorded with his own ideas and also fit most of the established principles of the time. Ultimately, this equation, and the physical theory behind it, proved to be one of the great intellectual achievements of the period. It was particularly remarkable for the internal beauty of its mathematical structure, which not only clarified previously mysterious phenomena such as spin and the FermiDirac statistics associated with it, but also predicted the existence of an electron-like particle of negative energy, the antielectron, or positron, and, more recently, it has come to playa . role of great importance in modem mathematics, particularly in the interrelations between topology, geometry, and analysis. Heisenberg characterized the discovery of antimatter by Dirac as "the most decisive discovery in connection with the properties or the nature of elementary particles .... This discovery of particles and antiparticles by Dirac. .. changed our whole outlook on atomic physics completely." One of the interesting implications of

168

6. GENERALIZED FUNCTIONS

his work that predicted thepositron was theprediction of a magnetic monopole. Dirac won the Nobel Prize in 1933 for this work. Dirac is not only one of the chief authors of quantum mechanics, buthe is also the creator of quantum electrodynamics andone of the principal architects of quantum field theory. While studying the scattering theory of quantum particles, he invented the (Dirac) delta junction; in his attempt at quantizing the geueral theory of relativity, he founded constrained Hamiltonian dynamics, which is one of the most active areas of theoretical physicsresearch today. One of his greatest contributions is theinvention of bra ( Iand ket

I)· Whileat Cambridge, Diracdidnot accept many research students. Thosewho worked with him generally thought that he was a good supervisor; but one who did not spend muchtimewithhis students. A student neededto be extremely independent to work under Dirac. One such student was Dennis Sciama, who later became the supervisor of Stephen Hawking, the current holder of the Lucasian chair. Salam and Wigner, in their preface to the Festschrift that honors Diraconhis seventieth birthday andcommemorates his contributions to quantum mechanics succinctly assessedtheman: Diracis one of the chief creators of quantum mechanics .... Posterity will rateDiracas one of thegreatest physicists of all time.Thepresent generation values him as one of its greatest teachers . . .. On those privileged to know him, Dirac has left his mark . .. by his human greatness. He is modest, affectionate, and setsthehighestpossiblestandards of personal and scientific integrity. He is a legendin his own lifetimeandrightly so. (Taken from Schweber, S. S. "Some chapters for a history of quantum field theory: 19381952", in Relativity, Groups, and Topology II vol. 2, B. S. Dewitt and R. Stora, eds., North-Holland, Amsterdam, 1984.)

We have seen that the delta functiott Cattbe thought ofas the limit ofatt ordinary

function, This idea can he generalized, 6.2.4. Defittition. Let {'Pn (x)} be a sequence offunctions such that

n~i: 'Pn(x)f(x)dx exists for all f E e~(JR). Then the sequence is said to converge to the distribution 'P, defined by ('P, f) =

j}!'!"J: 'Pn(x)f(x)dx

V f.

This convergence is denoted by 'Pn -> 'P. For example, it Catt be verified that

n 2 2 _ e - n x -> 8(x)

.,fii

attd

1 - cosnx -> 8(x)

nnx 2

and so Ott. The proofs are left as exercises.

6.3 PROBLEMS

derivative of a distribution

169

6.2.5. Definition. The derivative of a distribution rp is another distribution rp' defined by (rp', f) = - (rp, f') V f E e~. 6.2.6. Example. Wecancombine thelasttwo definitions to showthat if thefunctions en are defined as

a B,,(x)

==

if x < _1

(nx

{1

+ 1)/2

--k ~"~ s k. k,

if

if x ~

tbenB~(x) -+ 8(x).

Wewritethe definition of the derivative, (B~,

1

00

1

f) = -

1

00

, B,,(x)f(x)dx = -

00

df B,,(x)-dx = Bn(x)df -00 dx -00 1 l n /" Bn(X)d f+l / Bn(x)df+

-00

=_(1-

-00

= _

=

-lin

1/" nx + Idf + roo d f) (0 + 1-I/n 2 11/n

_::1

1n / xdf _

2 -I/n

=

(Bn • j'), in terms of integrals:

~

I

1 /"

df _

2 -I/n

-~ (xf(x)I~I~" - 2: (f(lfn) -

1

i:

roo Bn(X)d f)

11/n

roo df

11/n f(X)dX)

f(-I/n» - f(oo)

+ f(l/n).

Forlarge n. we have I/n '" 0 and f(±I/n) '" frO). Thus,

1

00

-00

, n (I I I I 2 ) B,,(x)f(x)dx '" - - - f(-) + - f(--) - -frO) 2nnnnn

The approximation becomesequality in the limit n -+ 00

lim

n"""*oo

1

,

Bn(x)f(x)dx = frO) = (80. f)

-00

Note that /(00) finite volume.

=

=>

00.

+ frO)

'" frO).

Thus,

,

Bn -+ 8.

0 because of the assumption that all functions must vanish outside a II

6.3 Problems 6.1. Write a density function for two point charges q; and qz located at r = rl and r = r2. respectively. 6.2. Write a density function for four point charges ql = q, q2 = -q, q3 = q and q4 = -q, located at the comers of a square of side 2a, lying in the xy-plane, whose center is at the origin and whose first comer is at (a. a).

170

6. GENERALIZED FUNCTIONS

I

6.3. Show that 8(f(x)) =

8(x - xo), where xo is a root of f and x is 1f'(xo)1 confined to values close to xo. Hint: Make a change of variable to Y = f(x).

6.4. Show that

where the

Xk'S

are all the roots of

f

in the interval on which f is defined.

e

6.5. Define the distribution p : oo (R3) --+ R by (p, f) =

f


S

where
6.6. Define thedisttibution p : eoo (R3) --+ R by (p, f) =

fc

)..(r)f(r) di(r) ,

where )..(r) is a smooth function on a smooth curve C in R 3 • Show that per) is zero if r is not 00 C and infinite if r is on C.

6.7. Express the three-dimensional Dirac delta function as a product of three onedimensional delta functions involving the coordinates in (a) cylindtical coordinates, (b) spherical coordinates, (c) general curvilinear coordinates. Hint: The Dirac delta function in R 3 satisfies JJJ 8 (r)d 3x = 1. 6.8. Show that J""oo 8'(x)f(x) dx = - f'(O) where 8'(x) sa !x8(x). 6.9. Evaluate the following integrals:

(a)

i:

8(x 2 - 5x

+ 6)(3x 2 -7x + 2) dx.

(e) roo 8(sin]fx)(~)Xdx.

10.5

(b)

i: 00

(d)

1

-00

8(x 2 - ]f2)cosxdx. 2

8(e- X )lnxdx.

Hint: Use the result of Problem 6.4. 6.10. Consider

IxI as a generalized function and find its derivative.

6.3 PROBLEMS

171

6.11. Let ~ E eOO(JRn) be a smooth function on JRn, and let
the sense of Definition6.2.4. n

2 2

(a) .,jiie-n x •

(b) 1- cosnx 1tnx 2

(d) sinnx. nx

Hint: Approximate
+ tanhnx) --> O(x) as n --> 00.

6.14. Show thatx8'(x) = -8(x).

Additional Reading 1. Hassani, S. Mathematical Methods, Springer-Verlag, 2000. An elementary treatment of the Dirac delta function with many examples drawn from mechanics and electromagnetism. 2. Rudin, W. Functional Analysis, McGraw-Hill, 1991. Part II of this mathematicalbut (forthosewith a strongundergraduatemathematicsbackground) very readable book is devotedto the theory of distributions. 3. Reed, M. and Simon, B. Functional Analysis, Academic Press, 1980.

7

_

Classical Orthogonal Polynomials

The last example of Chapter 5 discussed only one of the many types of the so-called classical orthogonal polynomials. Historically, these polynomials were discovered as solutions to differential equations arising in various physical problems. Such polynomials can be produced by starting with I, x, X 2, ... and employing the Gram-Schmidt process. However, there is a more elegant, albeit less general, approach that simultaneously studies most polynomials of interest to physicists. We will employ this approach.'

7.1 General Properties Most relevant properties of the polynomials of interest are contained in

7.1.1. Theorem. Consider the functions 1 dn Fn(x) = w(x) dx n (uis")

for n = 0, 1,2, ... ,

(7.1)

where 1. F, (x) is a first-degree polynomial in x, 2. s (x) is a polynomial in x of degree less than or equal to 2 with only real roots,

3. w(x) is a strictly positive function, integrable in the interval (a, b), that satisfies the boundary conditions w(a)s(a) = 0 = w(b)s(b). IThis approach is dueto F. G. Tricomi [Tric 55]. See also [Denn 67].

7.1 GENERAL PROPERTIES

173

Then Fn (x) is a polynomial of degree n in x and is orthogonal-on the interval (a, b), with weight w(x)---to any polynomial Pk(X) of degree k < n, i.e., Pk(x)Fn(x)w(x) dx = Ofor k < n. These polynomials are collectively called classical onhogonal polynomials.

J:

Before proving the theorem, we need two lemmasf 7.1.2. Lemma. The following identity holds: m:S n.

D

Proof See Problem 7.1.

7.1.3. Lemma. All the derivatives d/" jdxm(ws n) vanish at x = a and x = b,for all values ofm < n.

Proof Set k = 0 in the identity of the previous lemma and let P,o = 1. Then we dm have --(ws n) = ws n-mFsm- The RHS vanishes atx = a and x = b due to the dx'" third condition stated in the theorem. D

Proof of the theorem. .We prove the orthogonality first. The proof involves multiple use of integration by parts:

l

b

pk(x)F"(x)w(x)dx =

a

lb l

n

I d Pk(x)[-n (ws n)] wdx

a

W

dx

d [d n

b

1

Pk(X)- - --- I (uis") ] dx a dx dx»: b n- I b d I dPk d"-l = Pk(X) - - I (ws") d - d1 (ws n) dx. dxv: a a X x ll =

l

~

=0 by Lemma 7.1.3

This shows that each integration by parts transfers one differentiation from uis" to Pk and introduces a minus sign. Thus, after k integrations by parts, we get

lb

b k n k d Pk d a Pk(x)Fn(x)w(x)dx = (_I)k a dxk dxn_k(wsn)dx

l

d [d n-kn) ] = C a dx dx" k 1 (ws dx

l

b

I

2Recall that P5k is a generic polynomial with degree less than or equal to k.

b

d"-k-l n) I = C dx" k 1 (ws a = 0,

174

7. CLASSICAL ORTHOGONAL POLYNOMIALS

where we have used the fact that the kth derivative of a potynomial of degree k is a constant. Note that n - k - 1 ~ 0 because k < n, so that the last line of the equation is well-defined. To prove the first part of the theorem, we use Lemma 7.1.2 with k = 0 and m = dn 1 dn n to get -(wsn) = wP
l

l P~n_,Fn(x)w(X) b

b

[Fn(x)]2w(x)dx =

dx

+ kn

l

b

x n Fn(x)w(x)dx.

The LHS is a positive quantity because both w(x) and [Fn(x)]2 are positive, and the first iotegral on the RHS vaoishes by the first part of the proof. Therefore, the second term on the RHS cannot be zero. In particular, kn 0, and Fn (x) is of degreen. D

t=

It is customary to iotroduce a normalization constant io the definition of Fn (x), and write

1 dn Fn(x) = - - __ (ws n). Knwdx n

(7.2)

generalized Rodriguez formula

This equation is called the generalized Rodriguez fonnula. For historical reasons, different polynomial functions are normalized differently, which is why K. is introduced here. From Theorem 7.1.1 it is clear that the sequence Fo(x), F, (x), F2(X), . " of polynomials forms an orthogonal set of polynomials on [a, b] with weight function w(x). All the varieties of classical orthogonal polynomials were discovered as solutions of differential equations. Here, we give a siogle generic differential equation satisfied by all the Fn's. The proof is outlioed in Problem 7.4.

differential equation forclassical orthogonal polynomials

7.1.4. Proposition. Let k, be the coefficientofx in Fl (x) and (72 the coefficient of x 2 in sex). Then the orthogonal polynomials Fn satisfy the differential equation3

We shall study the differential equation above in the context of the SturmLiouville problem (see Chapters 18 and 19), which is an eigenvalue problem involving differential operators.

3A prime is a symbol forderivative withrespect to x.

175

7.2 CLASSIFICATION

7.2 Classification Let us now investigate the consequences of vatious choices of s(x). We start with Ft (x), and note that it satisfies Equation (7.2) with n = 1:

1 d Ft(x) = --(ws),

or

Kiui dx

which can be integrated to yield ws = Aexp(f KtFt(x)dx/s) where A is a constant. On the other hand, being a polynomial of degree 1, Fl (x) can be written as F;(x) = klX + k;. It follows that w(x)s(x) = Aexp

(I

Kl (klX

s

+ kJ) dx, )

w(a)s(a) = 0 = w(b)s(b).

(7.3)

Next we look at the three choices for s(x): a constant, a polynomial of degree 1, and a polynomial of degree 2. For a constant s(x), Equation (7.3) can be easily integrated: w(x)s(x) = Aexp

(I

Kl(klXs + k1 'dx ) ) =Aexp

= Aeax2+px+c =

(I

(2ax+fJ)dx

)

Beax2+px.

The interval (a, b) is determined by w(a)s(a) = 0 = w(b)s(b), which yields Beua2+pa = 0 = Beub2+Pb. The only way that this equality can hold is for a and b to be infinite. Since a < b, we must take a = -00 and b = +00, in which case a < O. With y = .JIiiT(x +fJ/(2a)) and choosing B = s exp(fJ2 /(4a)), we obtain w(y) = exp( _y2). We also take the constant sto be 1. This is always possible by a proper choice of constants such as B. lfthe degree of sis 1, then s(x) = O'IX + 0'0 and

where y = Klkl/O'I, P = Klk1JO'I - KlklO'O/O'[, and B is A modified by the constant of integration. The last equation above must satisfy the boundary conditions at a and b: B(O'la + O'o)Peya = 0 = B(O'lb + O'O)Peyb, which give a = -0'0/0'1, P > 0, Y < 0, and b = +00. With appropriate redefinition of vatiables and parameters, we can write w(y) = yVe-Y, v > -1, and s(x) = x, a = 0, b = +00.

176

7. CLASSICAL ORTHOGONAL POLYNOMIALS

fL

v

w(x)

Polynomial

0

0

I

Legendre, Pn(x)

J.-! J.-! 1

-2 1

2 Table 7.1

1

(I - x 2jA- l/2

Gegenbauer, C~ (x), J. >

-!

-2

(I - x 2 ) - 1/ 2

Chebyshev of the first kind, Tn(x)

1

(I - x 2 ) 1/2

Chebyshev of the second kind, Un(x)

2

Special casesof Jacobi polynomials

Similarly, we can obtain the weight function and the interval of integration for the case when s (x) is of degree 2. This result, as well as the results obtained above, are collected in the following proposition. 7.2.1. Proposition.

If the conditions ofTheorem 7.1.1 prevail,

then

(a) For s(x) ofdegree zero we get w(x) = e- x' with s(x) = I, a = -00, and b = +00. The resulting polynomials are called Hermite polynomials and are denoted by Hn (x). (b) For s(x) of degree 1, we obtain w(x) = xVe- x with v > -1, s(x) = x, a = 0, and b = +00. The resulting polynomials are called Laguerre polynomials and are denoted by L~(x).

(c) Fors(x) of degree 2, we get w(x) = (1 +x)I'(I-x)V with u.,» > -1, s(x) = 1 - x 2, a = -1, and b = +1. The resulting polynomials are called Jacobi polynomials and are denoted by p/:. v (x). Jacobi polynomials are themselves divided into other subcategories depending on the values of fL and v. The most common and widely used of these are collected in Table 7.1. Note that the definition ofeach ofthe preceding polynomials involves a "standardization," which boils down to a particular choice of K n in the generalized Rodriguez formula.

7.3 Recurrence Relations Besides the recurrence relations obtained in Section 5.2, we can use the differential equation of Proposition 7.1.4 to construct new recurrence relations involving derivatives. These relations apply only to classical orthogonal polynomials, and notto general ones. We start with Equation (5.12) Fn+l(X) = (anx

+ fJn)Fn(x) + )'nFn-l(X),

(7.4)

7.3 RECURRENCE RELATIONS

177

differentiate both sides twice, and substitnte for the secoud derivative from the differeutial equation of Propositiou 7.1.4. This will yield

2wsanF~ + [an:X (ws) + WAn(anX + fin)] t; - WAn+!Fn+1

+ WYnAn-IFn-l

(7.5)

= O.

Karl Gustav Jacob Jacobi (1804-1851) was the secoud son born 10 a well-to-do Jewish baoking family in Potsdam. An obviously bright young man, Jacobi was soon moved to the highest class in spite of his youth and remained at the gym-

nasium for four years only because he could not enter the university until he was sixteen. He excelled at the University of Berlin in all the classical subjects as well as mathematical

studies, the topic he soon chose as his career. He passed the examinationto become a secondaryschool teacher,then later the examination that allowed university teaching, and joined the faculty at Berlin at the age of twenty. Since promotion there appeared unlikely, he moved in 1826 to the University of Konigsberg in search of a more permanent position. He was known as a lively and creative lecturer who often injected his latest research topics into the lectures. He began what is now a common practice at most universities-the research seminar-for the most advanced students and his faculty collaborators. The Jacobi "school," together with the influence of Bessel and Neumann (also at Konigsberg), sparked a renewal of mathematical excellence in Germany. In 1843 Jacobi fell gravely ill with diabetes. After seeing his condition, Dirichlet, with the help of von Humboldt, secured a donation to enable Jacobi to spend several months in Italy, a therapy recommended by his doctor. The friendly atmosphere and healthful climate there soon improved his condition. Jacobi was later given royal permission to move from

Konigsberg 10 Berlin so that his heal!h would not be affected by the harsh winlers in the former location. A salary bonus given to Jacobi to offset the highercostofliving in the capital was revoked after he made some politically sensitive remarks in an impromptu speech. A permanent position at Berlin was also refused, and the reduced salary and lack of security caused considerable hardship for Jacobi and his family. Only after he accepted a position in Vienna did the Prussian government recognize the desirability of keeping the distinguished mathematician within its borders, offering him special concessions that together with his love for his homeland convinced Jacobi to stay. In 1851 Jacobi died after contracting both influenza and smallpox. Jacobi's mathematical reputation began largely with his heated competition with Abelin the study ofelliptic functions. Legendre, formerly the star of such studies, wrote Jacobi ofhis happiness at having "lived long enough to witness these magnanimous contests between two young athletes equally strong." Although Jacobi and Abel could reasonably be considered contemporary researchers who arrived at many of the same results independently, Jacobi suggested the names "Abelian functions" and "Abelian theorem" in a review he wrote for Crelle's Journal. Jacobi also extended his discoveries in elliptic functions to number theory and the theory of integration. He also worked in other areas of number theory, such as the theory ofquadratic forms and the representation of integers as sums of squares and cubes. He

178

7. CLASSICAL ORTHOGONAL POLYNOMIALS

presentedthewell-known Jacobian, orfunctional determinant, in 1841.Tophysicists, Jacobi is probably best known for his workin dynamics with the form introduced by Hamilton. Althoughelegant andquite general, Hamiltonian dynamics did not lend itselfto easysolution of many practical problems in mechanics. In the spirit of Lagrange, Poisson, and others, Jacobi investigated transformations of Hamilton's equations thatpreserved their canonical nature (loosely speaking, thatpreserved thePoisson brackets in eachrepresentation). After much work and a little simplification, the resulting equations of motion, now known as Hamilton-Jacobi equations, allowedJacobi to solve several important problems in ordinary andcelestial mechanics. Clebsch andlater Helmholtz amplified their use in other areas of physics.

We can get another recurrence relation involving derivatives by substituting (7.4) in (7.5) and simplifying:

ZwsanF~ + [an :X (ws) + wO'n - An+t)(anx + fJn)] Fn + wYn(An-l - An+l)Fn-l = O.

(7.6)

Two other recurrence relations can be obtained by differentiating Equations (7.6) and (7.5), respectively, and using the differential equation for Fn • Now solve the first equation so obtained for Yn(djdx)(wFn-l) and substitote the result in the second equation. After simplification, the result will be

ZwanAnFn + :x {[an :x (ws)

+ (An-I -

+ W(An -

An-l)(anx + fJn)] Fn} (7.7)

d

An+t) dx (WFn+l) = O.

Finally, we record one more useful recurrence relation:

dw An(x)Fn - An+l(anx + fJn)-d Fn+t + YnAn-l(anx x

+ Bn(x)F~+t + YnDn(x)F~_1 =

dw

+ fJn)-Fn-1 dx

0,

(7.8)

where

An (x) = (anx

+ fJn) [zwanA n + an

2 d 2 (ws) dx

+ An(anx + fJn) dW] dx

2 d - an dx (ws),

d Bn(x) = an dx (ws) - w(anx

Dn(x) = w(anx

+ fJn)(An-l

+ fJn)(An+1

- An),

d - An) - an -(ws). dx

Details of the derivation of this relation are left for the reader. All these recurrence relations seem to be very complicated. However, complexity is the price we pay

179

7.4 EXAMPLES OF CLASSICAL ORTHOGONAL POLYNOMIALS

useful recurrence relations for Hermite and Legendre polynomials

for generality. When we work with specific orthogonal polynomials, the equations simplify considerably. For instance, for Hermite and Legendre polynomials Equation (7.6) yields, respectively, (l - X2 )Pn'

and

+ nxp"

- nPn-l = O.

(7.9)

Also, applying Equation (7.7) to Legendre polynomials gives

,

,

Pn+1 - XPn - (n + I)Pn = 0,

(7.10)

and Equation (7.8) yields

,

,

Pn+ 1 - Pn-1 - (2n + I)Pn = O.

(7.11)

It is possible to find many more recurrence relations by manipulating the ex-

istingrecurrence relations. Before studying specific orthogonal polynomials, let us pause for a moment to appreciate the generality and elegance of the foregoing discussion. With a few assumptions and a single defining equation we have severely restricted the choice of the weight function and with it the choice of the interval (a, b). We have nevertheless exhausted the list of the so-called classical orthogonal polynomials.

7.4 Examples of Classical Orthogonal Polynomials We now consttuct the specific polynomials used frequently in physics. We have seen that the four parameters K n, k n, k~, and h n determine all the properties of the polynomials. Once K n is fixed by some standardization, we can detenmine all the other parameters: k n and k~ will be given by the generalized Rodriguez formula, and h n can be calculated as

hn =

l

b F;(x)w(x) dx

=

l

b

(knx n + ... )Fn(x)W(X) dx

l

b n-1 n d [d I d k wxn----(wsn)dx = -'0. x n_ - - I (ws n) ] dx a Knw dx" Kn a dx dx»: bd n-1 n- I k xn _ d _ (ws n) [b _ -'0. k d _ (ws n) dx = -'0. _(xn) _ Kn dxr:" a K n a dx dx n- 1

=kn

l

b

l

The first term of the last line is zero by Lemma 7.1.3. It is clear that each integration by parts introduces a minus sign and shifts one differentiation from uis" to x". Thus, after n integrations by parts and noting that dO j dx o( ws n) = ius" and d" jdxn(x n) = n!, we obtain

hn =

(-l)nknn! Kn

l

a

b WS

nd

x.

(7.12)

180

7. CLASSICAL ORTHOGONAL POLYNOMIALS

7.4.1 Hermite Polynomials summary of properties of Hermite polynomials

The Hermite polynomials are standardized such that K n = (_1)n. Thus, the geueralized Rodriguez formula (7.2) and Proposition 7.2.1 give 2 d" 2 Hn(x) = (_I)"e x - n (e- X ). dx

(7.13)

x2 It is clear that each time e- is differentiated, a factor of -2x is introduced. x2 The highest power of x is obtained when we differentiate e- n times. This yields 2 2 (_I)nex (_2x)n e - x 2nx n k n 2n. To obtain l(,., we find it helpful to see whether the polynomial is even or odd. We substitute -x for x in Equation (7.13) and get Hn(-x) = (_I)n Hn(x), which shows thatifn is even (odd), Hn is an even (odd) polynomial, i.e., it can have only even (odd) powers of x. In either case, the next-highest power of x in Hn(x) is not n - I but n - 2. Thus, the coefficient of x n- 1 is zero for Hn(x), and we have k~ = O. For h«, we use (7.12) to obtain h n = .,jii2nnL Next we calculate the recurrence relation of Equation (5.12). We can readily calculate the constants needed: an = 2, fin = 0, Yn = -2n. Then substitute these in Equation (5.12) to obtain

'*

=

=

(7.14) Other recurrence relations can be obtained similarly. Finally, the differential equation of Hn(x) is obtained by first noting that K 1 = -I, (jz = 0, Fl (x) = 2x kl = 2. All of this gives An = -2n, which can be used in the equation of Proposition 7.1.4 to get

'*

dZHn an; dx z - 2x dx

+ 2nHn =

(7.15)

O.

7.4.2 Laguerre Polynomials summary of properties of Laguerre polynomiais

For Laguerre polynomials, the standardization is K n = nL Thus, the generalized Rodriguez formula (7.2) and Proposition 7.2.1 give L~(x)

=

dn (xVe-xx n) n!xVe x dx" I

n

I d = _x-VeX_(xn+ve-X).

n!

dx"

(7.16)

To find k n we note that differentiating e-x does not introduce any new powers of x but ouly a factor of -1. Thus, the highest power of x is obtained by leaving x n+v alone and differentiating e- x n times. This gives

~,x-vexXn+V(_I)ne-X

n.

=

(_I)n xn n!

'* kn = (_I)n. n!

181

7.4 EXAMPLES OF CLASSICAL ORTHOGONAL POLYNOMIALS

We may try to check the evenness or oddness of L~ (x); however, this will not be helpful because changing x to -x distorts the RHS of Equation (7.16). In fact, k~ i' 0 in this case, and it can be calculated by noticing that the next-highest power of x is obtained by adding the first derivative of x"+v n times and multiplying the result by (_1)n-l , which comes from differentiating e- x. We obtain

I ( 1)"-I(n + v) ,,-1 -x -v eX[( - 1)"-1 n (n + v )x n+v-l e -X] = )1 X , n! (n - I . and therefore k~ = (-I)"-I(n Finally, for h" we get

h n -_ (-1)"[(-1)" /n!]n! n!

thegamma function

+ v)/(n -

1"" 0

I)L

x v e -x x "d x -_ -I n!

1"" 0

x n+v e-x d x.

If v is not an integer (and it need not be), the integral on the RHS cannot be evaluated by elementary methods. In fact, this integral occurs so frequently in mathematical applications that it is given a special name, the gamma function. A detailed discussion of this function can be found in Chapter 11. At this point, we simply note that

['en + 1) = n! for n EN,

(7.17)

and write h n as

h

n

= ['en + v + 1) = ['en + v + 1) . n!

['en + 1)

The relevant parameters for the recurrence relation can be easily calculated: Ol

1 n+I'

----

,,-

_ 2n+v+I fJn- n+l '

n+v Y"=-n+I'

Substituting these in Equation (5.12) and simplifying yields

(n + I)L~+l = (2n + v + I - x)L~ - (n + v)L~_I' With kl = -I and (f2 = 0, we get A" Proposition 7.1.4 becomes d2 p

x--f dx

ai:

+ (v + I-x)--" +nL~ = dx

= -n, and the differential equation of

O.

(7.18)

182

7. CLASSICAL ORTHOGONAL POLYNOMIALS

7.4.3 Legendre Polynomials summary of properties of Legendre polynomials

Instead of discussing the Jacobi polynomials as a whole, we will discuss a special case of them, the Legendre polynomials P" (x), which are more widely used in physics. With IJ. = 0 = v, corresponding to the Legendre polynomials, the weight function for the Jacobi polynomials reduces to w(x) = I. The standardization is K" = (- 1)"2"n!. Thus, the generalized Rodriguez formula reads

(_I)" a" 2nn ! dx"

2

P,,(x) = - - - - [(1- x)"].

(7.19)

To find k", we expand the expression in square brackets using the binomial theorem and take the nth derivative of the highest power of x. This yields

k x" _ (_I)" a" " - 2" n., ax II

[(

_x2)" _ _ 1_ a" (x2n) ] 2" n.I x

= _1_ 2n(2n - 1)(2n -

2nn !

a" 2) ... (n + I)x".

After some algebra (see Problem 7.7), we get k" =

2"r(n + 1) 1 2

n!r(z)

Adrien-Marie Legendre (1752-1833) came from a wellto-do Parisian family and received an excellent education in science and mathematics. His university work was advanced enough that his mentor used many of Legendre's essays in a treatise on mechanics. A man of modest fortune until the revolution, Legendre was able to devote himself to study and research without recourse to an academic position. In 1782 he won the prize of the Berlin Academy for calculating the trajectories of cannonballs taking air resistance into account. This essay brought him to the attention of Lagrange and helped pave the way to acceptance in French scientific circles, notably the Academy of Sciences, to which Legendre submitted numerous papers. In July 1784 he submitted a paper on planetary orbits that contained the now-famous Legendre polynomials, mentioning that Lagrange had been able to "present a more complete theory" in a recent paper by using Legendre's results. In the years that followed, Legendre concentrated his efforts in number theory, celestial mechanics, and the theory of elliptic functions. In addition, he was a prolific calculator, producing large tables of the values of special functions, and he also authored an elementary textbook _that remained in use for many decades. In 1824 Legendre refused to vote for the government's candidate for Institut National. Because of this, his pension was stopped and he died in poverty and in pain at the age of 80 after several years of failing health. Legendre produced a large number of useful ideas but did not always develop them in the most rigorous manner, claiming to hold the priority for an idea if he had presented

7.4 EXAMPLES OF CLASSICAL ORTHOGONAL POLYNOMIALS

183

merely a reasonable argument for it. Gauss, with whom he had several quarrels over priority. considered rigorous proof the standard of ownership. To Legendre's credit, however, he was an enthusiastic supporter of his young rivals Abel and Jacobi and gave their work considerable attention in his writings. Especially in the theory of elliptic functions, the area of competition with Abel and Jacobi, Legendre is considered more of a trailblazer than a great builder. Hermite wrote that Legendre "is considered the founder of the theory of elliptic functions" and "greatly smoothed the way for his successors," but Dotes that the recognition of the double periodicity of the inverse function, which allowed the great progress of others, was missing from Legendre's work. Legendre also contributed to practical efforts in science and mathematics. He and two of his contemporaries were assigned in 1787 to a panel conducting geodetic work in cooperation with the observatories at Paris and Greenwich. Four years later the same panel members were appointed as the Academy's commissioners to undertake the measurements and calculations necessary to determine the length of the standard meter. Legendre's seemingly tireless skill at calculating produced large tables of the values of trigonometric and elliptic functions, logarithms, and solutions to various special equations. In his famous textbook Elements de geometrie (1794) he gave a simple proof that 7r is irrational and conjectured that it is not the root of any algebraic equation of finite degree with rational coefficients. The textbook was somewhat dogmatic in its presentation of ordinary Euclidean thought and includes none of the non-Euclidean ideas beginning to be fanned around that time. It was Legendre who first gave a rigorous proof of the theorem (assuming all of Euclid's postulates, of course) that the sum of the angles of a triangle is "equal to two right angles." Very little of his research in this area was of memorable quality. The

same could possibly be arguedfor the balanceof his writing,bot one must acknowledge the very fruitful ideas he left behind in number theory and elliptic functions and, of course, the introduction of Legendre polynomials and the important Legendre transformation used both in thennodynamics and Hamiltonian mechanics.

To find k~, we look at the evenness or oddness of the polynomials. By an investigation of the Rodriguez formula-s-as in our study ofHennite polynomialswe note that Fn( -x) = (_I)n Fn(x), which tells us that Fn(x) is either even or odd. In either case, x will not have an (n - l)st power. Therefore, k~ = O. We now calculate h n as given by (7.12): h

n

= (-I)nknn! s;

1 1

-1

(l _ x 2)ndx

=2

nr(n

+ ~)/ r(~) 2nn !

1 1

(1 _ x 2)ndx. _1

The integral can be evaluated by repeated integration by parts (see Problem 7.8). Substituting the result in the expression above yields h n = 2/(2n + 1). We need ctn , fin and Yn for the recurrence relation: n

ct =

k n+ l

k::

2n + 1r(n + 1 +~) n!r(~) 2n + 1 = (n + 1)!r(~) 2nr(n +~) = n + 1 '

where we used the relation I'(n-l- 1 +~) = (n+ ~)r(n+ ~). We also have fin = 0 (because ej, = 0 = k~+I) and Yn = -n/(n+l). Therefore, the recurrence relation

184

7. CLASSICAL ORTHOGONAL POLYNOMIALS

is

(n

+ l)Pn+I(X) =

(2n

+ l)xPn(x) -nPn-I(X).

(7.20)

Now we use K, = -2, Pl(X) = x =} kl = 1, and 02 = -1 to obtain An = -n(n + 1), which yields the following differential equation:

d [ 2 dPn] dx (l-x) dx = -n(n + l)Pn·

(7.21)

This can also be expressed as

ar;

2P 2 d n (l - x ) dx 2 - 2x dx

+ n(n + l)Pn = O.

(7.22)

7.4.4 Other Classical Orthogonal Polynomials The rest of the classical orthogonal polynomials can be constructed similarly. For the sake of completeness, we merely quote the results.

Jacobi Polynomials, P::'v (x) K; = (-2)nn!

Standardization: Constants:

hn =

kn

= 2-n r(2n + fJ. + v + 1), n!r(n + fJ.

k'

=

+ v + 1) n 2,,+v+lr(n + fJ. + l)r(n + v + 1) -:-:c::-----'----'--'---=-'----'-------'----'-,.,n!(2n + fJ. + v + l)r(n + fJ. + v + 1)

n(v - fJ.) k 2n + fJ. + v n,

Rodriguez formula:

Differential Equation: d2 P"'v dx 2

dP"'v v - (fJ. + v + 2)x]_n_ dx + n(n + fJ. + v + l)P/:'v = 0

(1 - x 2) _ _ n-

+ [fJ. -

A Recurrence Relation: 2(n

+ l)(n + fJ. + v + 1)(2n + fJ. + v)P:-i-~ = (2n + fJ. + v + 1)[(2n + fJ. + v)(2n + u. + v + 2)x + v 2 - 2(n + fJ.)(n + v)(2n + fJ. + v + 2)P:~~

fJ.2]p/:. v

7.4 EXAMPLES OF CLASSICAL ORTHOGONAL POLYNOMIALS

Gegenbauer Polynomials, C;(x) f(n + A + ~)r(2A) K - (_2)nn'-'-_ _---"'--'-..,..:n . r(n + 2A)f(A + ~)

Standardization:

Constants:

_ 2n f(n + A) k n - n! f(A) ,

,

~,

..jiif(n + 2A)r(A + ~) - 0 h - --'--;c;--:-:-=-:;;:-:-,-=,-!'-, n - n!(n + A)r(2A)r(A)

''"n -

Rodriguez Formnla:

Differential Equation:

d2e'" . de'" - (2A + l)x d" + n(n + 2A)e~ =

(1 - x 2 )-----f dx

x

0

A Recurrence Relation:

(n + l)e~+l = 2(n

+ A)Xe~ -

(n + 2A - l)e~_l

Chebyshev Polynomials of the First Kind, Tn(x)

Standardization:

K = (_I)" (2n)! n 2 nn!

Constants:

• (_I)n2nn! dn . (l_x 2)1/2_ [(l_x 2)" - 1/2] Rodrfguez Formulaz Tn(x) = (2n)! dx" Differential Equation:

A Recnrrence Relation:

2

d 2Tn dx

dTn dx

(1 - x ) - - 2 - x -

+ n 2 Tn =

0

Tn+l = 2xTn - Tn-l

Chebyshev Polynomials of the Second Kind, U« (x) Standardization: Constants:

K _ _ 1)n (2n + I)! n-( 2n(n+l)!

185

186

7. CLASSICAL ORTHOGONAL POLYNOMIALS

Rodriguez Formula: n(n U ( ) = (_1)n2 + 1)1 (1 _ x 2)- 1/2 d n [(1 _ x 2)" +1/ 2] n x (2n + I)! dx" 2

d 2Un

au;

dx

dx

Differential Equation: (1 - x )-2- - 3 x - + n(n A Recurrence Relation:

7.5

+ 2) Un = 0

Un+1 = 2xUn - Un-I

Expansion in Terms of Orthogonal Polynomials

Having studied the different classical orthogonal polynomials, we can now use them to write an arbitrary function f E £'~(a, b) as a series of these polynomials. If we denote a complete set of orthogonal (not necessarily classical) polynomials by ICk) and the given function by If), we may write 00

(7.23)

If) = Lak ICk), k=O

where ak is found by multiplying both sides of the equation by (Cil and using the orthogonality of the ICk) 's: 00

(Cil f) = Lad Cil Ck) = a; (Cil Cj) k=O

=} aj

=

(Cil f) . (CilCj)

(7.24)

This is written in function form as

t

C~(x)f(x)w(x) dx a· - ",a'-:-i:--'-'-=--.:-'--'-:-'---'---'----i

r-:

J;ICj(x)1 2w(x)dx

(7.25)

We can also "derive" the functional form of Equation (7.23) by multiplying both ofits sides by (xl and using the fact that (xl f) = f(x) and (xl Ck) = Ck(X). The result will be 00

f(x) = LakCk(X). k=O

(7.26)

7.5.1. Example. The solution of Laplace's equation in spherically symmetric electro-

static problems that are independent of theazimuthal angleis givenby

~( HI bk (r, 0) = L.. k=O r

+ ckr k) Pk(COSO).

(7.27)

Consider two conducting hemispheres of radius a separated by a small insulating gap at theequator. Theupper hemisphere is heldat potential Vo andthelowerone at - Vo. as

7.5 EXPANSION IN TERMS OF ORTHOGONAL POLYNOMIALS

187

shown in Figure 7.1. We want to find the potential at points outsidethe resulting sphere. Since the potential mustvanishat infinity, we expectthe second term in Equation (7.27) to be absent, i.e., Ck = 0 V k. To find bb substitute a for r in (7.27) and let cosB '" x. Then, bk HI Pk(X),

00

(a, x) = L k=Oa

where (a, x) =

V. - 0 { +Vo

- I < x < 0,

if

ue

if

x -c L

From Equation (7.25), we have 2k + 111 -2Pk(X) (a, x) dx -1

Toproceed, we rewrite thefirst integral:

1

0 Pk(x)dx

-1

=_

fO

Pk(-y)dy

1+1

=

f1 Pk(-y)dy = (_I)k f1 Pk(x)dx,

10

10

where we madeuse oftbe parity property of Pk(X). Therefore, bk

2k+1

k

a H 1 = -2-Vo[l- (-I) ]

r

10

Pk(x)dx.

It is now clear that only oddpolynomials contribute to the expansion. Using the result of Problem 7.26, we get

or bZm+1

= (4m + 3)a Zm+2VO(-I) m 2Zm +1(2m)! . ml(m + 1)1

Note that (a, x) is an odd function; that is, (a, -x) = -(a, x) as is evident from its definition. Thus,only oddpolynomials appear in the expansion of 4>(a,x) to preserve this property. Havingfoundthe coefficients, we can writethe potential: 00

(r,B)=VoL(-I) k~O

mZ(4m+1+ 3)(2m)1 .(a)zm+z PZm+1(cos B). 2 m ml(m + 1)1

r

III

188

7. CLASSICAL ORTHOGONAL POLYNOMIALS

Figure 7.1 Thevoltageis +Vo for theupper hemisphere, where0 :::; B < 1T/2, orwhere cosO:::; 1. It is -Vo for the lower hemisphere, where 7C/2 < B :::; rr , or where -1 ~cose <0.

o<

The place where Legendre polynomials appear most natnrally is, as mentioned above, in the solntion of Laplace's eqnation in spherical coordinates. After the partial differential equation is transformed into three ordinary differential equations using the method of the separation of variables, the differential equation corresponding to the polar angle e gives rise to solutions of which Legendre polynontials are special cases. This differential equation simplifies to Legendre differential equation if the substitution x = cos is made; in that case, the solutions will be Legendre polynontials in x, or in cos e. That is why the argument of Pk(X) is restricted to the interval [-I, +I].

e

7.5.2. Example. We can expand the Dirac delta function in tenus of Legendre polynomials.Wewrite 00

8(x) =

L a"P,,(x),

where

an

2n+11 =-

n=O

2

1

-1

Pn(x)8(x) dx

=

2n+1 --Pn(O). 2

For odd n this will give zero, because Pn (x) is an oddpolynomial. To evaluate Pn (0) for even n, we use therecurrence relation(7.20) for x = 0:

or nPn(O) obtain

= -en -1)P,,_z(O), or P,,(O) =

n-I ---Pn-z(O). Iteratingthis m times, we n

R 0 _ I m (n - I)(n - 3)··· (n - 2m + I) ,,( ) - (-) n(n - 2)(n - 4) ... (n - 2m + 2) P,,-Zm(O). (2m - 1)(2m - 3) .. ·3' I PoCO). Nowwe "fill 2m(2m - 2) .. ·4· 2 the gaps" in the numerator by multiplying it-and the denominator, of course-by the

For n = 2m, this yields PZm(0) = (_i)m

7.6 GENERATING FUNCTIONS

189

denominator. Thisyields P2m (0) = (_I)m

2m(2m -1)(2m - 2)· .. 3·2·1 [2m(2m _ 2) ... 4 . 2]2

m (2m)! = (-I)

because Po(x) ~(x)

m

[2mm!]2 = (-I)

(2m)! 2 2m(m!)2'

1. Thus, we canwrite

=

~ 4m + I m = L.--(-I)

2

m=O

(2m)! 2 2 P2m(x). 2 m(m!)

We can also derive this expansion as follows. For any complete set of orthonormal vectors [IA}}~t, we have ~(x - x')

= w(x) (xl x') = w(x) (xI1Ix ') = w(x) {xl

(~llk) (AI) Ix'} = w(x) ~ Ik(x')lk(X).

Legendre polynomials are not orthonormal; hut we can makethem so by dividing Pk(x) by h~/2 = ..j2/(2k + I). Then,notingthat w(x) = I, we obtain

~(x -

,

x)

00

Pk(X')

= ~ ..j2((2k + I)

For x' = 0 we get ~(x) = result.

Pk(X) ..j2(2k + I)

00 2k + I , = ~ -2-Pk(x )Pk(X).

E~o 2k; 11'k(O)Pk(X),

which agrees with the previons III

7.6 Generating Functions generating function

It is possible to generate all orthogonal polynomials of a certain kind from a single function of two variables g(x, t) by repeated differentiation of that function. Such a function is called a generating function. This generating function is assumed to be expandable in the form 00

g(x, t) = I>ntn Fn(x),

(7.28)

n=O

so that the nth derivative of g(x, t) with respect to t evaluated at t = 0 gives Fn(x) to within a multiplicative constant. The constant an is introduced for convenience. Clearly, for g (x, t) to be useful, it must be in closed form. The derivation of such a function for general Fn(x) is nontrivial, and we shall not attempt to derive such a general generating function-as we did, for instance, for the general Rodriguez formula. Instead, we simply quote these functions in Table 7.2, and leave the derivation of the generating functions of Hermite and Legendre polynomials as Problems 7.14 and 7.20. For the derivation of Laguerre generating function, see [Hassani, 2000] pp. 606-607.

190

7. CLASSICAL ORTHOGONAL POLYNOMIALS

Polynomial

Generating function

an

exp(-t 2 + 2xt)

lin!

exp[-xtl(l-t)]/(l-t)v+l

I

Hermite, Hn(x) Lagnerre,L~(x)

Chebyshev (1st kind), Tn(x)

(I - t 2 )(t2 - 2xt

+ 1)-1

2, n

f= 0

ao = 1 Chebyshev (2nd kind), Un (x)

Table 72

(t 2

-

2xt + 1)-1

1

Generating functions for selected polynomials

7.7 Problems 7.1. Let n = 1 in Eqnation (7.1) and solve for s dw . Now snbstitute this in the dx derivative of ws nP,k and show that the derivative is eqnalto ws n-l P,k+!. Repeat this process m times to prove Lennna 7.1.2. 7.2. Find w(x), a, and b for the case of the classical orthogonal polynomials in which s(x) is of second degree. 7.3. Integrate by parts twice and nse Lennna 7.1.2 to show that

.\

l

b

Fm(wsF~)' dx =0

for m < n.

7.4. (a) Using Lemma 7.1.2 conclnde that (wsF~)' [u: is a polynomial of degree less than or equal to n. (b) Write (ws F~)'Iw as a linear combination of F, (x), and use their orthogonality and Problem 7.3 to show that the linear combination collapses to a single term. (c) Multiply both sides ofthe differential equation so obtained by Fn and integrate. The RHS becomes hnAn. For the LHS, carry out the differentiation and note that (ws)' I w = Kl Fl. Now show that K 1 Fl F~ + s F;: is a polynomial of degree n, and that the LHS of the differential equation yields {(Klkln + 02n(n - I)}h n. Now find An. 7.5. Derive the recurrence relation of Equation (7.8). Hint: Differentiate Equation (7.5) and substitute for from the differential equation. Now multiply the resulting equation by anx + fJn and substitute for (anx + fJn)F~ from one of the earlier

F;:

recurrence relations.

7.7 PROBLEMS

191

7.6. Using only the orthogonality of Hermite polynomials

i:

e-X'Hm(x)Hn(x)dx = .j1iZnn!8 mn

(and the fact that they are polynomials) generate the first three of them.

7.7. Show that for Legendre polynomials, kn = Znr(n + !)/[n!r(~)]. Hint: Multiply and divide the expression given in the book by n!; take a factor of Z out of all terms in the nmnerator; the even terms yield a factor of n!, and the odd terms give a gannna function. 7.8. Using integration by parts several times, show that 1 (1 _ x2)"dx = zmn(n - I)··· (n - m + I) -I . 3 . 5 . 7 ... (Zm - 1)

1

Now show that I~I (1 - x 2)ndx = Zr(!)n!/[(Zn

II

x2m(l _ x2)n-mdx.

-I

+ 1)r(n + !)].

7.9. Use the generalized Rodriguez formula for Hermite polynomials and integration by parts to expand x 2k and x 2k+I in terms of Hermite polynomials. 7.10. Use the recurrence relation for Hermite polynomials to show that 00

1

-00 xe-x

,

Hm(x)Hn(x)dx = .j1i2n- In! [8m •n -

1

+ 2(n + 1)8m •n +l ] '

What happens when m = n?

7.11. Apply the general forma1ism of the recurrence relations given in the book to Hermite polynomials to find the following:

Hn + H~_I - 2xHn_ 1 = O. 7.12. Show that 1':"00 x 2e-x ' H;(x) dx = .j1i 2n(n

+ !)n!

7.13. Use a recurrence relations for Hermite polynomials to show that

Hn(O) =

{O (_I)m (2~)!

ifn is odd,

ifn=2m.

7.14. Differentiate the expansion of g (x, t) for Hermite polynomials with respect to x (treating t as a constant) and choose an such that nan = an-I to obtain a differential equation for g. Solve this differential equation. To determine the "constant" ofintegration use the result ofProblem 7.13 to showthatg(O, t) = e-/'.

192

7. CLASSICAL ORTHOGONAL POLYNOMIALS

7.15. Use the expansion of the generating function for Hermite polynontials to obtain

Then integrate both sides over x and use the orthogonality of the Herntite polynontials to get (sz)" fOO 2 L: -,-Z e- H;(x)dx = ./iieZ". n~O (n.) 00

X

-00

Deduce from this the normalization constant hn of Hn (x). 7.16. Using the recurrence relation of Equation (7.14) repeatedly, show that

f

OO

k _x 2

Xe

Hm(x)Hm+n(x)dx =

{Orz: '171:

-00

if n > k,

2m(m

+ k)!

if n = k.

7.17. Given that Po(x) = I and PI (x) = x, (a) use (7.20) repeatedly to show that Pn(l) = 1. (b) Using the same equation, find Pz(x), P3(X), and P4(X). 7.18. Apply the general formalism of the recurrence relations given in the book to find the following , , two relations for Legendre polynontials: (a) nPn - xP; + Pn- 1 = O. (b) (I - xZ)P~ - nPn-l + nXPn = O. 7.19. Show that J~l x n Pn(x) dx = 2n+1(n!)z j(2n + I)!. Hint: Use the definition of hn and kn and the fact that Pn is orthogonal to any polynontial of degree lower thann.

7.20. Differentiate the expansion of g (x, t) for Legendre polynontials, and choose = I. For P~, you will substitute two different expressions to get two equations. First use Equation (7.11) with n + I replaced by n, to obtain Un

(I -

d 00 tZ)~ +tg = 2 L:ntn Pn-l + 2t. dx

n=2

As an alternative, use Equation (7.10) to substilnte for

dg

00

dx

n=Z

P~

and get

(I-xt)- = L:ntnPn_l +t.

Combine the last two equations to get (lz - 2xt + I)g' = tg. Solve this differential equation and detemtine the constant of integration by evaluating g (x, 0).

7.7 PROBLEMS

193

7.21. Use the generating function for Legendre polynontials to show that Pn (I) = I, Pn(-I) = (_I)n, Pn(O) = 0 for odd n, and P~(l) = n(n + 1)/2. 7.22. Both electrostatic and gravitational potential energies depend on the qnantity I/lr - r'], where r' is the position of the sonrce (charge or mass) and r is the observation point. (a) Let r lie along the z-axis, and nse spherical coordinates and the definition of generating functions to show that _1_ = ~ f(r<)n Pn(cose), [r - r'] r-; n=O r-;

e

where r «r» is the smaller (larger) of rand r', and is the polar angle. (b) The electrostatic or gravitational potential energy (r) is given by (r) =

kll!

') p(r d 3x ' , [r - r']

where k is a constant and p(r') is the (charge or mass) density function. Use the result of part (a) to show that if the density depends only on r', and not on any angle (i.e., p is spherically symmetric), then (r) reduces to the potential energy of a point charge at the origin for r > r', (c) What is (r)-in the form of an integral-for r < a for a spherically symmetric density that extends from origin to a? (d) Show that E (or g) is given by [kQ(r)lr 2]er where Q(r) is the charge (or mass) enclosed in a sphere of radius r, 7.23. Use the generating function for Legendre polynontials and their orthogonality to derive the relation

1 1

1

1 dx 2 = f t 2n P;(x)dx. _11-2xt+t n=O -1

Integrate the LHS, expand the result in powers of t, and compare these powers on both sides to obtain the normalization constant h n • 7.24. Evaluate the following integrals using the expansion of the generating function for Legendre polynontials. (a)

"

+ b) sine de . 2+2abcose+b2 o -v'a (a cos s

10

7.25. Differentiate the expansion of the Legendre polynontial generating function with respect to x and manipulate the resulting expression to obtain 00

00

(I - 2xt + t ) :~:::>n P~(x) = t ~::>n Pn(x). 2

n=O

n=O

194

7. CLASSICAL ORTHOGONAL POLYNOMIALS

Equate equal powers of t ou both sides to derive the recurrence relation

,

,

,

Pn+ 1 + Pn- 1 - 2xPn - Pn = O. 7.26. Show that

[1 Pk(X) dx

io

=

if k is even,

kO {8 (_I)(k-l'/2(k_l)!

S",==:fu2-" k 2 (y

if k is odd.

)!( &¥ )!

Hint: For even k, extend the region of integration to (-I, I) and use the orthogonality property. For odd k, note that dk - 1 2 k 1 dx k - 1 (I-x) 10

gives zero for the upper limit (by Lemma 7.1.3). For the lower limit, expand the expression using the binomial theorem, and carry out the differentiation, keeping in mind that only one term of the expansion contributes.

parity relations

7.27. Showthatg(x, r) = g(-x, -t)forbothHermiteandLegendrepolynomiais. Now expand g(x, t) and g(-x, -t) and compare the coefficients of t" to obtain the parity relations for these polynomials: Pn(-x) = (_I)n P,,(x).

and

7.28. Derive the orthogonality of Legendre polynomials directly from the differential equation they satisfy. 7.29. Expand Ixl in the interval (-I, +1) in terms of Legendre polynomials. Hint: Use the result of Problem 7.26. 7.30. Apply the general formalism of the recurrence relations given in the book to find the following two relations for Laguerre polynomials: v

v

(a) nLn - (n + v)L n_ 1 (b) (n

+ I)L~+1 -

(2n

-

dL~

x dx = O.

+v + I -

x)L~

+ (n + v)L~_1 =

O.

7.31. From the generating function for Laguerre polynomials given in Table 7.2 deduce that L~(O) = r(n + v + I)/[n!r(v + I)]. 7.32. Let L n sa L2. Now differentiate both sides of e-xt/(l-t)

g(x,t)=

I-t

00

= LtnL,,(x)

o with respect to x and compare powers of t to obtain L~ (0) = -n and L~ (0) = ~n(n - I). Hint: Differentiate 1/(1 - t) = I:;;:o t" to get an expression for (I - t)-2.

7.7 PROBLEMS

195

7.33. Expande-kx as a series ofLaguerre polynomials L~ (x). Find the coefficients by using (a) the orthogonality of L~(x) and (b) the generating function. 7.34. Derive the recurrence relations given in the book for Jacobi, Gegenbauer, and Chebyshev polynomials. 7.35. Show that Tn(-x) g(x, t) = g(-x, -t).

= (-I)nTn(x)

and Un(-x)

= (-I)nun(x). Hint: Use

7.36. Show that Tn(l) = I, Un(l) = n + I, Tn(-I) = (_I)", Un(-I) = (-I)n(n + I), T2m(O) = (_l)m = U2m(O), and T2m+l(O) = 0 = U2m+! (0).

Additional Reading J. Dennery, P. and Krzywicki, A. Mathematicsfor Physicists, Harper and Row, 1967. Treats the classical orthogonal polynomials in the spirit of this chapter. 2. Tricomi, F. Vorlesungen uber Orthogonalreihen, Springer, 1955. The original unified treatment of the classical orthogonal polynomials.

8,

_

Fourier Analysis

The single most recurring theme of mathematical physics is Fourier analysis. It shows up, for example, in classical mechanics and the analysis of normal modes, in electromagnetic theory and the frequency analysis of waves, in noise considerations and thermal physics, in quantum theory and the transformation between momentum and coordinate representations, and in relativistic quantum field theory and creation and annihilation operation formalism.

8.1 Fourier Series One way to begin the study of Fourier series and transforms is to invoke a generalization of the Stone-Weierstrass Approximation Theorem (Theorem 5.2.3), which established the completeness of monomials, x k . The generalization of Theorem 5.2.3 permits us to find another set of orthogonal functions in terms of which we can expand an arbitrary function. This generalization involves polynomials in more than one variable (For a proof of this theorem, see Sinunons [Sinun 83, pp 160-161].) generalized Stone-Weierstrass theorem

8.1.1. Theorem. (generalized Stone-Weierstrass theorem) Suppose that !(X\, X2, ... , x n) is continuous in the domain {ai ::: Xi ::: hi li=\. Then it can be expanded in terms ofthe monomials X~l X~2 ••. x~n , where the k; are nonnegative integers. Now let us consider functions that are periodic and investigate their expansion in terms of elementary periodic functions. We use the generalized StoneWeierstrass theorem with two variables, x and y. A function g(x, y) can be written as g(x, y) = I:k,m~o akmxkym. In this equation, x and y can be considered as coordinates in the xy-plane, which in tum can be written in terms ofpolar coordinates

8.1 FOURIER SERIES

r and

197

e. In that case, we obtain 00

[(r, e) sa g(r cos s, r sine) =

e sin" e.

akmrk+m cosk

L k,m=O

e

In particular, if we let r = I, we obtain a function of alone, which upon substitution of complex exponentials for sin and cos becomes

e

e

(8.1) where b" is a constant that depends on akm' The RHS of (8.1) is periodic with period 2IT; thus, it is especially suitable for periodic functions [(e) that satisfy the periodicity condition [(e - IT) = [(e + IT). We can also write Equation (8.1) as 00

[(e) = bo + L (b"ei"e + b_"e-i"e) n=l 00

= bo + L[(b"

+ b_,,) cosne + i(b"

n=l ' " - r - - '

- b_,,) sinne)]

., =Bn

=A n

I

00

= bo + L(A" cosne

+ B" sinne).

(8.2)

n=l

If [(e) is real, then bo, A", and B" are also real. Equation (8.1) or (8.2) is called the Fourier series expansion of [(e). Let us now concentrate on the elementary periodic functions ei"e. We define the (Ie,,) }~I such that their "eth components" are given by

(el e,,) =

1 .e l'C:e"', -v2IT

where

e E (-IT, IT).

These functions-or ket vectors-which belong to .(,2(-IT, IT), are orthonormal, as can be easily verified. It can also be shown that they are complete. In fact, for functions that are continuous on (-IT, IT), this is a result of the generalized Stone-Weierstrass theorem. It turns out, however, that (Ie,,) }~I is also a complete orthonormal sequence for piecewise continuous functions on (-IT, IT).I Therefore, any periodic piecewise continuous function of e can be expressed as a linear combination of these orthonormal vectors. Thus if If) .(,2(-IT, IT), then

e

00

If) =

L [" Ie,,),

where

f" =

(e,,1 f).

(8.3)

n=-oo

1A piecewisecontinuous function onafiniteinterval is one that hasa finitenumber of discontinuities in itsinterval ofdefinition.

198

8. FOURIER ANALYSIS

Fourier series expansion: anguiar

We can write this as a functional relation if we take the Oth component of both sides: (01 f) = '£':::-00 fn (01 en), or

expression

f(O) = _1_

f

fnein9

(8.4)

.[iii n=-oo with fn given by fn

= (enI 1 If) = (enl (j~ 10)(01 dO) If) = =

-1-1"

L:

e-in9 f(O)dO.

(enl 0) (01 f)

so (8.5)

.[iii -"

It is important to note that even though

f

(0) may be defined only for -:n: :0:

o :0: :n:, Eqnation (8.4) extends the domain of definition of f(O) to all the intervals

fundamental cell ofa periodic function

(2k - I):n: :0: 0 :0: (2k + I):n: for all k E Z. Thns, if a function is to be represented by Equation (8.4) without any specificationof the interval of definition, it must be periodic in O. For such functions, the interval of their definition can be translated by a factor of2:n:. Thus, f(O) with -:n: :0: 0 :0: :n: is equivalentto f(O - 2m:n:) with 2m:n: -:n: :0: 0 :0: 2m:n: +:n:; both will give the same Fourier series expansion. We shall define periodic functions in their fundamental cell such as (-:n:, :n:).

"The profound study JosephFourier (1768-1830) didvery wellasayoung student ofnature isthe most of mathematics but had set his heart on becoming an army fruitful source of officer. Denied a commission because he was the son of a mathematical tailor, he went to a Benedictine school with the hope that he discoveries."

Joseph Fourier

couldcontinue studying mathematics at its seminary in Paris. TheFrench Revolution changed thoseplans andset the stage formany of the personal circumstances ofFourier's lateryears, due in part to his courageous defense of some of its victims, an action that led to his arrest in 1794. He was released later that year, andhe enrolled as a student in theEcole Nonnale, which opened and closedwithin ayear. Hisperformance there, however, wasenough toeamhimaposition asassistant lecturer (under Lagrange and Monge) in the Ecole Polytechnique. He was an excellentmathematical physicist, was a friend of Napoleon (sofarassuchpeoplehavefriends), andaccompanied himin 1798toEgypt, where Fourier held various diplomatic andadministrative posts while also conducting research. Napoleon took note of his accomplishments and, on Fourier's return to France in 1801, appointed him prefect of thedistrict of Isere, in southeastern France, and in this capacity builtthe first realroad from Grenoble to Turin. He also befriended the boy Champollion, who later deciphered the Rosetta stone as the first long step toward understanding the hieroglyphic writing of the ancient Egyptians. Like other scientists of his time, Fourier took up the flow of heat. The flow was of interest as a practical problem in the handling of metals in industry and as a scientific problem in attempts to determine thetemperature in theinterior of the earth, the variation

8.1 FOURIER SERIES

199

of that temperature with time, and other such questions. He submitted a basic paper on heat conduction to the Academy of Sciences of Paris in 1807. The paper was judged by Lagrange, Laplace, and Legendre. The paper was not publish
Functions are not always defined on (-17:,17:). Let us consider a function F(x) that is defined on (a, b) and is periodic with period L = b - a. We define a new variable,

and note that f(O)

f(U17:)

Fourier series expansion: general expression

== F«LI217:)O + a + L12) has period (-17:,17:) because

= F (2~ (0 ±1r)+a +~) = F (x ±~)

and F(x + L12) = F(x - LI2). If follows that we can expand the latter as in Equation (8.4). Using that equation, but writing 0 in tenus of x, we obtain

200

8. FOURIER ANALYSIS

F(X)=F(~8+a+£)=_I. 2:n:

f

2 . . / i i i n=-oo

!n exp [in

2:n: L

(x-a-£)] 2

f

= _1_ Fne2mrixjL, 4 n=-oo

(8.6)

where we have introduced- Fn sa '/L/2:n:!ne-i(2nn/L)(a+L/2). Using Equation (8.5), we can write

t; =

[Le- i(2nn/L)(a+L/2)_I_ In e- inO!(8)d8 ../iii -n a L = 4 e-i(2nn/L)(a+L/2) + e-i(2nn/L)(x-a-L/2) F(x) 2:n: dx ~ a L b = _1_1 e- i(2nn/L)xF(x)dx. 4a

V2;.

l

(8.7)

The functions exp(2:n:inx/L)/4 are easily seen to be orthonormal as members of .c}(a, b). We can introduce {len)}~1 with the "xth component" giveu by (xl en) = (1/4)eZrrinx/L. Then the reader may check that Equations (8.6) and (8.7) can be written as IF) L~-oo Fn len) with Fn (nl F).

=

=

8.1.2. Example. In the study of electrical circuits, periodic voltage signals of different square wave voltage

shapes are encountered. An example is a square wave voltage of height Uo."duration" T, and "rest dnration" T [see Fignre 8.I(a)]. The potential as a function of time Vet) can be expanded as a Fourier series. Theinterval is (0, 2T) because that is one whole cycle of the potential variation. Wetherefore use Equation (8.6) and write Vet)

I

= --

L 00

Vn,e2n1fitj2T,

where

I Jo2T e-2n,nt/2T . Vn = - _ V(t)dt.

..tiT

..tiTn=-oo

0

Theproblem is to find Vn . Thisis easily done by substituting V t _ {Uo if 0 s t ::; T, ()- 0 if T s t s 2T

in the last integral:

Vn = Uo fT e-innt/T dt ..tiT 10

_ {O -

= Uo (_~) [(_I)n _

ifn is even

..tiTuo . 1n1f

..tiT

I] where n i' 0

1n1f

andn =1= 0,

if n IS . odd .

2The Fn aredefined suchthatwhattheymultiply in theexpansion areorthonormal in the interval (a, b).

8.1 FOURIER SERIES

+

u0 "- ..••....+........... :::::::::::t::::::::::: •••••• · ••• l ••••• ··•····

::::::::::::1:::::::::::

! (10............ .

.

~

:~:~:~:~:r: : :~:~:~I: : : : : : L: : : : :I:~:~:~:~: ~:L:~:~:~:~:l: : : : : :I: : :.:.~

o -.::.-

U

.••..••••••• ..:............. • •••••••••••

o

:

--.-:

.:10............

--:;;.:

;

:

~

..

-T

(a)

.

,~l~'.~~H~;f~~~:~·~~~~~ -2T

201

(b)

2T

T

Figure 8.1 (a) The periodic square wave potential. (b) Various approximations to the Fourier series of the square-wave potential. The dashed plot is that of thefirst term of the series, the thick grey plot keeps 3 terms, and the solid plot IS terms.

For n

= 0, we obtain Vo = "'"' I

1

2T

..,2T 0

we can write

V(t) = _1_ [uo

,.j2f

= Uo

{! 2

= "'"' I

V(t) dt

I

T

..,2T 0

2

Z:Tr

~ If: l1r

_I e-in"t/T

n=l -n

+

odd

{!

= Uo~ -. Therefore,

f'f + ,fiT Uo ( n=-oo I: !en imrt/ T + n=lf: !e'nm/T)] n

V"2

n odd

+

Uodt

n odd

f:

!e inm/ T]}

n=l n n odd

1ft)}.

~ ~ _1_ in ([2k+ 11 o 2+1ft;Q2k+l s T

= U

Figure 8.1(b)showsthegraphical representation of the above sumwhenonlyafinite number of tenus arepresent. ..

sawtooth voltage

8.1.3. Example. Another frequently nsed voltage is the sawtooth voltage [see Figure 8.2(a)]. The equatiou for V(t) with period T is Vet) = UotlT for 0 and its Fourier representation is

V(t)

=-

I

L 00

Jf n~-oo

Vne2n1titjT,

where

Vn = _1_

~ t ~

rT e-2"int/T V(t) dt.

Jf io

T,

202

B. FOURIER ANALYSIS

Substituting for V(t) in the integral above yields Vn = _1_ {T e-21rintjTUo!.... dt = UoT-3j2 {T e-21rint/T t dt ..ti T 0

10

= ti

1

r-3/2(---J!--e-2Jrint/TIT +;!.-

o

-12mr

(T e-21rint/T dt)

l2mr 10

0

::..::.-~-~o

2) T uo..ti = UOT- 3/2 ( - .= -.--12m/: z2n:tr Vo =

-1

leT V(t) dt =

..ti 0

1f'i'

-rt

where n t" 0,

leT UO-t dt = !Uo..ti. T

0

Thus,

(t

~ei2n"'/T +

V(t) = _1_ [lUo..ti _ uo..ti ...tf 2 z2rr n=-oo n

= Uo

f: ~ei2n"t/T)]

n=l n

{~- 2. f: ~Sin(2mrt)}. 2

JCn=ln

T

Figure 8.2(b)showsthe graphical representation of the above series keeping the first few ternM.



The foregoing examples indicate an important fact about Fourier series. At points of discontinuity (for example, t = T in the preceding two examples), the value of the functiou is not defined, but the Fourier series expansion assigns it a value-the average of the two values on the right and left of the discontinuity. For instance, when we substitute t = T in the series of Example 8.1.3, all the sine terms vanish and we obtain V(T) = Uo/2, the average of Uo (on the left) and 0 (on the rigbt). We express this as

V(T) = ![V(T - 0)+ V(T

+ 0)]

sa

! .....0 lim [V(T -

E)+ V(T

+ E)].

This is a general property of Fourier series. In fact, the main theorem of Fourier series, which follows, incorporates this property. (For a proof of this theorem, see [Cour 62].)

8.1.4. Theorem. The Fourierseries ofafunction f(O) that is piecewise continuous in the interval (-Jr, Jr) converges to

![f(O + 0) + f(O - 0)]

for - x < 0 <

![f(Jr) + f(-Jr)]

for 0 = ±Jr.

it;

Although we used exponential functions to find the Fourier expansion of the

two examples above, it is more convenient to start with the trigonometric series

B.l FOURIER SERIES

203

(b)

time

Figure 8.2 (a)Theperiodic saw-tooth potential. (b)Various approximations to the Fourier series of the sawtooth potential. The dashed plot is thatof the first term of the series,the thick grey plot keeps 3 terms, and the solid plot 15 terms.

when the expansion of a real function is songht. Equation (8.2) already gives such an expansion. All we need to do now is find expressions for An and Bn. From the definitions of An and the relation between bn and In we get

An

1

= bn + b-n = v2re !'C(fn + I-n) = - 1

I (-

f" e-ine1(0) so + - 1 f" eine 1(0) dO)

v'2ir v'2ir -" v'2ir -" ine = -1 [e+eineJ/(O)dO = -1 cosnOI(O) dO. 211:

f"

tt

-1t

L -11:

(8.8)

Similarly,

B n = -I

:n:

bo =

L _"

sinnOI(O) dO,

1 I !'Clo = -2

V 2n

f" 1(0) se

es

I -2 Ao.

(8.9)

1!-1C

So, for a function 1(0) defined in (-re, :n:), the Fourier trigonometric series is as in Equation (8.2) with the coefficients given by Equations (8.8) and (8.9). For a function F(x), defined on (a, b), the trigonometric series becomes

1 F(x) = -AD 2

~( Znn x . 2n:n:x) + L.. Ancos-- + Bn sm-- , n=1

L

L

(8.10)

204

8. FOURIER ANALYSIS

where

t

Cn;X)

An =

~

En =

~ (b sin (2nrrx) F(x) dx.

cos

Lin

F(x) dx ; (8.11)

L

A convenient rule to remember is that for even (odd) functions-which are necessarily defined on a symmetric interval around the origin-only cosine (sine) terms appear in the Fourier expansion.

8.1.5. Example. An alternating current is tnmed into a direct current by starting with a signal of the form V(t) ex: I sin lllt I. i.e., a harmonic function thatis nevernegative, as shown in Figure 8.3(a). Thenby proper electronics, one smooths out the "bumps" so that the output signal is very nearly a direct voltage. Let us Fourier-analyze the above signal. Since Vet) is even for -lC -c tot < 1C, we expect only cosine tenus to be present. Ifforthe timebeing we use e instead of cot,we canwrite I sin BI = Ao + L~l An cos n6, where

!

An = -1

1"

IsinOlcosnOdO=2

10"

nO

]'(-Jr

sinOcos nOdO

2 [(-I)n+ = -21o"t ~[sin(n + 1)0 - sin(n -I)O]dO = - - 2 - 11:0 n-l1! _{-~(+) 1r n -1 o

-

for e even and e

1]

ee D,

forn odd,

and Ao = (1/rr) J':." Isin0l dO = 4/rr. The expansionthen yields . 2 4 ~ cos2kwt Ismwtl=--- L., 2 ' tt k~l 4k - 1

rr

where in the sum we substituted 2k for n, and OJt for B. Figure 8.3(b) shows the graph of the series above when only the first few termsarekept. II

It is useful In have a representation of the Dirac delta function in terms of the present orthonormal basis of Fourier expansion. First we note that we can represent the delta function in terms of a series in any set of orthonormal functions (see Problem 8.23):

8(x - x') =

L fn(x)t,;'(x')w(x).

(8.12)

n

Next we use the basis of the Fourier expansion for which w(x) = 1. We then obtain 00

8(x -x') =

L n=-oo

e21r:inx/L e-2ninx'/L

~

E n=-oo

e 2n in(x-x')/ L .

205

8.1 FOURIER SERIES

(a)

a

(b) ..........~-'-'-~ ..........~.........~........"

oO:"""~.L.....~..L.,..~u.....~

a

-rc/ro

21C/ro

1C/ro

31C/ro

time

Figure8.3 (a) The periodic "monopolar" sine potential. (b) Various approximations to theFourier seriesof the "monopolar' sinepotential. Thedashed plotis that of thefirst term of the series, the thick grey plot keeps 3 terms, and the solid plot 15 terms.

8.1.1 The Gibbs Phenomenon

Gibbs phenomenon

The plot of the Fourier series expansions in Figures 8.1(b) and 8.2(b) exhibit a feature that is common to all such expansions: At the discontinuity of the periodic function, the truncated Fourier series overestimates the actual function. This is called the Gibbs phenomenon, and is the subject of this subsection. Let us approximate the infiuite series with a fiuite sum. Then fN(O)

~ .0 L.. fn eln

1 = --

v'2rr n=- N

1 (2n = 27C In dO' f(O')

o

1 = --

~ .0 1 L.. e ln - -

v'2rr n=- N '

v'2rr

fo2n" e- . 0' f(O') dO' In

0

N

L

ein(O-O'),

n=-N

where we substituted Equation (8.5) in the sum and, without loss of generality, changed the interval of integration from (-:n:, IT) to (0, 2:n:). Problem 8.2 shows that the sum in the last equation is

~

in(O-O') _ sin[(N + ~)(O - 0')]

L.. e

-

n=-N

1

sin['i(O - 0')]

It follows that fN(O)

=...!:.-

r: s«

2" Jo

f(O') sin[(N

+ ~)(O -

sin[~(O

0')] - 0')]

206

6. FOURIER ANALYSIS

= -

i:

I

2n:

sin[(N + l)tP] I j2n~0 . 1 2 sa d4>f(tP s1O(2tP) 2n: -0

dtPf(tP + 0)

-0

+ O)S(tP)·

(8.13)

~

=SCq,)

We wantto investigate the behavior of f N at a discontinuity of f. By translating the limits of integration if uecessary, we can assume that the discontinuity of f occurs at a point a such that 0 01 a 01 2n:. Let us denote the jump at this discontinuity for the function itself by !'>f, and for its finite Foutier sum by !'>fN:

!'>f sa f(a

+ f) -

f(a - f),

Then, we have !'>fN I

=2n

c:

dtPf(tP + a

+ f)S(tP)

-a-e

- -

j2n-u+
I

211: -a+€

t:

f)S(tP)

{j-u+< dtPf(tP + a + f)S(tP) + dtPf(tP + a + f)S(tP)} -a-e - - I {j2n-u-
c:

2rr

-a+€

2rr-a-e

+ -I

i:

2n -a+e

maximum overshoot inGibbs phenomenon calculated

dtP[f(tP + a

+ f) -

f(tP

+a -

}

f)]S(tP)

The first two integrals give zero because of the small ranges of integration and the continuity of the integrands in those intervals. The integrand of the third integral is almost zero for all values of the range of integration except when tP '" O.Hence, we can confine the integration to the small interval (-8, +8) for which the difference in the square brackets is simply ts], It now follows that

!'>f j~ sin[(N + !)tP] !'>f fo~ sin[(N + !)tP] !'>fN(8) '" -2 . 1 dtP '" 1 dtP, it _~ s1O(2 tP) n: 0 2 tP wherewe have emphasized the dependence of fN on 8 and approximated the sine in the denominator by its argument, a good approximation due to the smallness of tP. The reader may find the plot of the integrand in Fignre 6.2, where it is shown clearly that the major contribution to the integral comes from the interval [0, n:1(N + !)], where n:I(N +!) is the first zero of the integrand. Furthermore, it the result of the integral will is clear that if the upper limit is larger than n: I (N + decrease, because in each interval of length 2n:, the area below the horizontal axis is larger than that above. Therefore, if we are interested in the maximum overshoot

!),

8.1 FOURIER SERIES

207

of the finite sum, we must set the upperlimit equal to re/(N + ~). It follows firstly that the maximum overshoot of the finite sum occurs at re I(N + ~) RJ n t N to the right of the discontinuity. Secondly, the amount of the maximum overshoot is =X

~

I

(l1fN )m",

2!:>.f 1,,/(N+2) sin[(N + ~)
RJ -

2 = -tV re

1 1C

d
A.

0

tt

'I'

sinx -dx

0

x

RJ

(8.14)

1.179I1f.

Thus

8.1.6. Box. (Gibbs phenomenon) The finite (large-N) sum approximation ofthe discontinuous function overshoots the function itselfat a discontinuity by about 18 percent.

8.1.2 Fourier Series in Higher Dimensions It is instructive to generalize the Fourier series to more than one dimension. This generalization is especially useful in crystallography and solid-state physics, which deal with three-dimensional periodic structures. To generalize to N dimensions, we first consider a special case in which an N-dimensional periodic function is a product of N one-dimensional periodic functions. That is, we take the N functions f(J)(x) =

~

f:

ff iJe 2i" kx / Lj , j = 1,2, ... , N,

yLj k=-oo

and multiply them on both sides to obtain (8.15)

where we have used the following new notations: F(r) ea f(l)(xj)f(2)(X2)" .j
k",(kl,k2, ... .k»). gk = 2re (kilL!, ... , kNILN),

V=LI L2··· LN, Fk sa fkl ... fkN' r = (Xl, X2, ... , XN).

We take Equation (8.15) as the definition of the Fourier series for any periodic function of N variables (not just the product of N functions of a single variable). However, application of (8.15) requires some clarification. In one dimension, the

208

8. FOURIER ANALYSIS

shape of the smallest region of periodicity is unique. I! is simply a liue segment of length L, for example. In two and more dimensious, however, such regions may

havea variety of shapes. Forinstance, in two dimensions, they can be rectangles,

Wigner-Seitz cell

peutagons, hexagons, and so forth. Thus, we let V in Equation (8.15) stand for a primitive cell of the N -dimensional lattice. This cell is important iu solid-state physics, and (iu three dimeusions) is called the Wigner-Seitz cell. I! is customary to absorb the factor 1!.,fV iuto Fk, and write

L Fk eigk"

F(r) =

(8.16)

k

where the integral is over a siugle Wigner-Seitz cell. Recall that F(r) is a periodic function ofr. This means that when r is changed by R, where R is a vector describiug the boundaries of a cell, then we should get the same function: F(r + R) = F(r). When substituted iu (8.16), this yields F(r + R) Lk Fkeig•. (,+R) Lk e ig•. R Fkeig •." which is equal to F(r) if

=

=

(8.17)

reciprocallallice vectors

In three dimensions R = mI8t + m28z + m3a3, where mI, mz. and m3 are iutegers and 31, 3Z, and 33 are crystal axes, which are not generally orthogonal. On the other hand, gk = nlbl + nzhz + n3b3, where nl, nz, and n3 are iutegers, and bl, bz, and b3 are the reclprocal lattice vectors defined by

The reader may verify that bi . 3 j = 2:n:8ij. Thus

~. R =

(tnibi) . (tmj3j) = 1=1

1=1

~nimjbi' 3j I,l

3

= 2:n: Lmjnj = 2:n:(iuteger), j=l

and Equation (8.17) is satisfied.

8.2 The Fourier Transform The Fourier series representation of F (x) is valid for the entire realliue as long as F(x) is periodic. However, most functions encountered in physical applications are defined in some iuterval (a, b) without repetition beyond that iuterval. I! wonld be useful if we conld also expand such functions iu some form of Fourier "series." One way to do this is to star! with the periodic series and then let the period go to iufinity while extending the domain of the definition of the function. As a

8.2 THE FOURIER TRANSFORM

209

f(x)

L

)1001

a

b=a+L

(a)

a-L

a+L

a

a+2L

(b) Figure 8.4

(a) The function we want to represent. (b) The Fourier series representation

of thefunction.

specific case, suppose we are iuterested in representiug a function f(x) that is defined only for the interval (a, b) and is assigned the value zero everywhere else [see Figure 8.4(a)]. To begiu with, we might try the Fourier series representation, but this will produce a repetition of our function. This situation is depicted iu Figure 8.4(b). Next we may try a function gA (x) defined in the interval (a - A/2, b + A/2), where A is an arbitrary positive number:

o gA(X)=

f(x)

{o

if a - A/2 < x < a, if a-c x -c b,

b < x < b + A/2.

if

This function, which is depicted iu Figure 8.5, has the Fourier series representation gA(X) =

I

00

.,jL + A n=-oo

where gA.n =

L

I

l

gA,ne2hmxI(L+Al,

b A/ 2

+

.,jL + A a-A/2

. e-2l1rnxI(L+AlgA(X)dx.

(8.18)

(8.19)

We have managed to separate various copies of the original periodic function by A. It should be clear that if A ---> 00, we can completely isolate the function

210

8. FOURIER ANALYSIS

and stop the repetition. Let us investigate the behavior of Equations (8.18) and (8.19) as A grows without bound. First, we notice that the quantity k n defined by k n es 2n71:/ (L + A) and appearing in the exponent becomes almost continuous. In other words, as n changes by one unit, k n changes only slightly. This suggests that the terms in the sum in Equation (8.18) can be lumped together in j intervals of width Sn j , giving

where kj ea 2nj7l:/(L + A), and gA(kj) A) /271:]!'1k j in the above sum, we obtain

Fourier integral transforms

==

gA,nj' Substituting !'1nj = [(L

+

where we introduced gA(kj) defined by gA(kj) sa .J(L + A)/271: gA(k j). It is now clear that the preceding sum approaches an integral in the limit that A --+ 00. In the same limit, gA (x) --+ f (x), and we have f(x) =

~1°O

'V

21l'

f(k)eikXdk,

(8.20)

-00

where . . JL+A f(k) sa lim u(kj) = lun - - gA(kj) A--+oo

J+ oo

= lim A ....

=

L

_1_1 ...J2ir

A-+oo

A

271:.JL

I

+A

2n

lb+

Af

2 e-'kjXgA(x)dx .

a-A/2

00

f(x)e-ikxdx.

(8.21)

-00

Equations (8.20) and (8.21) are called the Fourier integral transforms of f(k) and f(x), respectively. 8.2.1, Example. Let us evaluatethe Fourier transform of the function defiuedby f(x) = {b if Ixl < a, D if [x] > a

(see Figure 8.6). From (8.21) we bave J(k) = _1_

."fiii

foo f(x)e-ikxdx = -00

_b_

."fiii

fa-a e-ikxdx

= 2ab

."fiii

whichis the function encountered (anddepicted) in Example 6.1.2.

(Siuka), ka

8.2 THE FOURIER TRANSFORM

a-N2 Figure 8.5

211

b+N2

By introducing theparameter A, we havemanaged to separate the copies of the function.

Letus discussthisresultin detail. First, notethatif a a constant function overthe entire real line, andwe get _ I(k)

=

2b

sinka lim - 21r e-e-oc k

= v

=

~ 00, thefunction

f (x) becomes

2b =7C8(k) '"21r

by the resultof Example 6.1.2. This is the Fourier transform of an everywhere-constant function (see Problem 8.12). Next, let b --+ 00 anda --+ 0 in sucha way that2ab, whichis the area under I(x), is I. Then I(x) will approach the delta function, and j(k) becomes _ I(k)

I

2ab sinka

I

sinka

= b.... lim - - - = lim - - = - . oo.j'iii ka .j'iii a.... O ka .j'iii a-+O

Heisenberg uncertainty relalion

So the Fourier transform of the delta function is the constant 1/~. Finally, we note that the width of I(x) is Sx = 2a, and the width of j(k) is roughly thedistance, on the k-axis,betweenits first two roots, k+ andk_. on either side of k = 0: t>k = k+ -k_ = 27C[a. Thus increasing the width of I(x) results in a decrease in the width of !(k). In other words, whenthe function is wide, its Fourier transform is narrow. In the limit ofinfinite width (a constant function), we get infinite sharpness (the deltafunction). The last two statements are very general. In fact, it can be shownthat Ax Ak ~ 1 for any function I (x). When both sides of this inequality are multiplied by the (reduced) Planck constant a == h/(23t}, theresultis the celebrated Heisenberg uncertainty relation:3 t>xt>p 2: Ii, where p = Ilk is the momeutum of the particle. Having obtained the transform of I(x), we

canwrite I I( x}=--

1

00

sinka _ b -2b- - eikxdk -

.j'iii -oo.j'iii

k

7C

1

00

-00

sinka - e ikxdk . k

III

3Inthe contextof the uncertainty relation, the widthof the function-the so-calledwave packet-measures theuncertainty in theposition x of a quantum mechanical particle. Similarly, the widthof the Fourier transform measures the uncertainty in k, whichis related to momentum p via p = hk,

212

B. FOURIER ANALYSIS

f(x)

b

ax _a Figure8.6

x

a

The square "bump" function.

8.2.2. Example. Let us evaluate the Fourier transform of a Gaussian g(x) = ae- bx2 witha,b > 0:

1 ../2ii

g(k) = -a-

00

e-b(x 2+ikxjb)dx = ae-

k2 4b /

../2ii

-00

1

00

e- b(x+ikj2b)2dx .

-00

Toevaluate thisintegral rigorously, we would haveto use techniques developed in complex analysis, whicharenotintroduced until Chapter 10 (see Example 10.3.8).However, we can ignore the factthat theexponent is complex, substitute y = x + ik/(2b), and write

1

00

e-b[x+ik/(2b)]2 dx =

-00

1

00

e-by2 dy

=

-00

E,

Vb

::"'e-k'j(4b), which is also a Gaussian.

Thus, we have g(k) =

-n»

We note again that the width of g(x), which is proportional to lj../b, is in inverse relation to the width of g(k), which is proportional to ../b. We thus have Ax Ak - 1. III Equations (8.20) and (8.21) are reciprocals of one another. However, it is not obvious that they are consistent. In other words, if we substitute (8.20) in the RHS of (8.21), do we get an identity? Let's try this:

_1_1 [_1_1 ..;z;r ..;z;r ~1 1 00

Ilk) =

dx e-ikx

-00

00

=

21l'

00

!(k')eik'Xdk']

-00

00

dx

-00

!(k')ei(k'-k)Xdk'.

-00

We now change the order of the two integrations:

1

00

Ilk) =

-00

dk' /(k')

[~1°O dx ei(k'-k)X] . 21t

-00

But the expression in the square brackets is the delta function (see Example 6.1.2). Thus, we have Ilk) = J~oo dk' !(k')8(k' - k), which is an identity.

213

8.2 THE FOURIER TRANSFORM

f

As in the case of Fourier series, Equations (8.20) and (8.21) are valid even if and! are piecewise continuous. In that case the Fourier transforms are written

as

![f(x

+ 0) + f(x -

0)] =

![!(k+O)+!(k-O)] =

~

'1271:

1

00

!(k)eikxdk,

-00

~1°O

'1271:

f(x)e-ikxdx,

-00

(8.22)

where each zero on the LHS is an E that has gone to its limit. It is useful to generalize Fourier transform equations to more than one dimension. The generalization is straightforward:

f(r) = !(k) =

1 (271:)n/2 I

(271: )n/2

f f

dnke'' k -r f(k)

'

dnxf(r)e-''k .r.

(8.23)

Let us now use the abstract notation of Chapter 6 to get more insight into the preceding results. In the language of Chapter 6, Equation (8.20) can be written as (8.24) where we have defined

(xlk) =

I 'k =e' x. '1271:

(8.25)

Equation (8.24) suggests the identification If) es If) as well as the identity 1=

L:

(8.26)

Ik) (kl dk,

which is the same as (6.1). Equation (6.3) yields (8.27)

(kl k') = 8(k - k'),

which upon the insertion of a unit operator gives an integral representation of the delta function:

8(k - k')

= (klllk') = (kl =

1

00

-00

(L:

Ix) (xl dX) Ik')

1

00

(kl x) (xl k') dx = - I dxei(k'-k)x. 21f -co

Obviously, we can also write 8(x - x') = [1/(271:)] J":'oo dkei(x-x'lk.

214

8. FOURIER ANALYSIS

If more than one dimension is involved, we use 8(k - k') = _1_ (211")n 8(r - r') = _1_ (211")n

f f

dnxei(k-k'j.r

'

dnkei(r-r'j.k

(8.28)

'

with the inner product relations

(r I k)

I

= (211")n/2 e

rk-r

1

-ik.r

{k] r) = (211")n/2 e

,

(8.29)



Equations (8.28) and (8.29) and the identification If) == If) exhibit a striking resemblance between [r) and [k), In fact, any given abstract vector If) can be expressed either in terms of its r representation, (r] f) = f(r), or in terms of its k representation, (kl f) == f(k). These two representations are completely equivalent, and there is a one-to-one correspondence between the two, given by Equation (8.23). The representation that is used in practice is dictated by the physical application. In quantum mechartics, for instance, most of the time the r representation, corresponding to the position, is used, because then the operator equations turn into differential equations that are normally linear and easier to solve than the corresponding equations in the k representation, which is related to the momentum.

Yukawa potential

8.2.3. Example. In thisexample we evaluate theFourier transform ofthe Coulomb potential VCr) of a point charge q: VCr) = «tr- TheFouriertransfonn is important in scattering experiments withatoms, molecules,andsolids.As we shallsee in thefollowing, theFourier transform of V (r) is notdefined. However, if we work with the Yukawapotential, qe-ar

Va(r)

= --, r

a> 0,

the Fourier transform will be well-defined,andwe can take the limit a Coulomb potential. Thus, we seek the Fourier transform of Va(r). We are working in threedimensionsandtherefore may write Va(k) =

I

(211")3/2

Iff

~

0 to recoverthe

d3xe-ik.r~. -ar

r

It is clear from the presence of r that spherical coordinates are appropriate. We are free

to pick anydirection as the z-axis. A simplifying choice in this case is the direction of k. So, we let k = Iklez = kez, or k . r = kr cos e, where e is the polar angle in spherical coordinates. Now we have Va(k)

= __ q_ (211")3/2

roo r 2dr {1C sln s se {21r dcpe-ikrcos(Je-ar.

10

10

10

r

The cp integration is trivial and gives 217'. The 0 integration is donenext:

{1C sinOe-ikrcos(J dO =

fo

11

-'-1

e-ikrudu =

~(eikr _ ikr

e-ikr).

8.2 THE FOURIER TRANSFORM

215

We thus have

Va(k)

= q(2n)

rX! drrze-ar _1_ Ci kr _ e-ikr)

1

r

(211:)3/2 0 = __ q_ ~

(211:)1/2 ik q

[00

10

ikr

dr [e(-a+iklr _ e-(a+ik)rj 00

I (e(-a+ik)r

= (211:)1/2

ik

00

1

-a + ik 0

+

e-(a+ik)r 1 ,,+ ik 0

)

.

Note how the factor e- ar has tamed the divergent behavior of the exponential at r -+ 00. This was the reason for introducing it in the first place. Simplifying the last expression yields Va(k) = (2ql../'iii)(~ + ~)-1. The parameter" is a measure of the range of the potential. It is clearthat the larger ex is, the smallerthe range. In fact, it was in response to the short range of nuclear forces that Yukawa introduced a. Forelectromagnetism, where therange is infinite, ex becomes zeroandVa(r) reduces to V (r). Thus, theFourier transform of the Coulomb potential is _ 2q I VCoul (k) = ../'iii k2 . If a charge distribution is involved, the Fourier transform will be different.

.

8.2.4. Example. The example above deals with the electrostatic potential of a point charge. Let us now consider the case where the charge is distributed overa finite volume. Thenthepotential is V( ) = r

fff

qp(r') d 3 , '"

Ir' _ r]

x

q

f

per') d 3 ,

II'"' _ r]

x ,

where qp(r') is the charge density at r'. andwe have used a single integral because d 3x' already indicates thenumber of integrations to be performed. Notethatwe havenormalized p (1'"') so that its integral over the volmne is I. Figore 8.7 shows the geometry of the situation. Making a changeof variables, R == r' - r, or r' = R + r, and d 3x' = d 3X, with R", (X, Y, Z), we get V(k) =

(211:~3/2f d 3xe-ik.rq f

r P(RR+ ) d 3X .

(8.30)

Toevaluate Equation (8.30), we substitute for peR + r) in terms of its Fourier transform,

peR + r) = __1_ fd3k'p(k')eik'.(R+r).

(8.31)

(211:)3/2

Combining (8.30) and (8.31), we obtain V(k)

= .s.. f (211:)3

ik'·R

d 3xd3X d 3k'_e__ p(k')i r.(k'-k) R ik'·R

=q fd 3X d 3k'_e -R

-(k') p

(_1_ f (211:)3

d 3xeir.(k'-kl)

-...::.:.....:.....:.~--~ ~(k'-k)

(8.32)

216

8. FOURIER ANALYSIS

dq

o Figure 8.7 TheFourier transform of thepotential of a continuous charge distribution at P is calculated usingthis geometry. What is nice about thisresult is that thecontribution of the charge distribution, p(k), has beencompletely factored out.Theintegral, asidefroma constant anda change in the sign of k, is simply the Fourier transform of the Coulomb potential of a point cbarge obtained in theprevious example. We cantherefore writeEquation (8.32) as V(k)

form factor

Fourier transform and the discovery of quarks

= (2".)

3/2

P(k)VCoul(-k)

=

4".qp(k)

Ikl 2



This equation is important in analyzing the structure of atomic particles. The Fourier transform ii (k) is directly measurable in scattering experiments. In a typical experiment a (cbarged) target is probed with a charged point particle (electron). If the analysis of the scattering data showsadeviation from 1/ k2 in thebehavior of V(k), theniteanbeconcluded that the target particle has a charge distribntion. More specifically, a plot of k 2iT (k) versus k givesthevariation of p(k:), theform factor, withk.1f theresulting graph is a constant, then p(k) is a constant, and the target is a point particle [p(k) is a constant for point particles, where p(r') ex 8(r - r')]. If there is any deviation from a constant function, p(k) mnst have a dependence on k, andcorrespondingly, thetarget particle musthavea charge distribution. The above discussion, when generalized to four-dimensional relativistic space-time, was the basis for a strong argument in favor of the existence of point-like particlesqnarks-inside a proton in 1968, when the results of the scattering of high-energy electrons off protons at the Stanford Linear Accelerator Center revealed deviation from a constant ~~~oofu~~~



8.2.1 Fourier Transforms and Derivatives The Fourier transform is very useful for solving differential equations. This is because the derivative operator iu r space turns into ordinary multiplication in k

8.2 THE FOURIER TRANSFORM

217

space. For example, if we differentiate f (r) in Equation (8.23) with respect to x iwe obtain

~ fer)

=

aXj

f (2rr~n/2 f I

(2rr)n/2

=

dnk~ei(klXl+.+kjxj+ ..+k"x")J()<) aXj

dnkCikj)e

ik.r

J(k).

That is, every time we differentiate with respect to any component of r, the corresponding component of k "comes down." Thus, the n-dimensional gradient is V fer) = (2rr)-n/2 dnkCik)e ik.r J()<), and the n-dimensional Laplacian is v 2 fer) = (2rr)-n/2 f d"k(_k 2)eik.r J(k). We shall use Fourier transforms extensively in solving differential equations later in the book. Here, we can illustrate the above points with a simple example. Consider the ordinary second-order differential equation

f

d 2y C2 dx 2

dy

+ Ci dx + CoY =

f(x),

where Co, Cl, and C2 are constaots. We can "solve" this equation by simply substituting the following in it: y(X) = 2y d

Jzn f f ./'iii

dkYCk)e

= __ 1_

dx?

ikx,

dky(k)k 2eikx,

.

f(x) =

Jzn f Jzn f

dkJ(k)e

+ Co)e ikx =

Jzn f

dkJ(k)e

:~

=

dky(k)(ik)e

ikx,

ikx.

This gives

Jzn f

dky(k)(-C2 k2 + iCt k

ih

Equating the coefficients of e ikx on both sides, we obtain - k _ J(k) y( ) - -C2k2 + i Cik: + Co'

If we know J(k) [which can be obtained from f(x)], we can calculate y(x) by Fourier-transforming Y(k). The resulting integrals are not generally easy to evaluate. In some cases the methods of complex analysis may be helpful; in others numerical integration may be the last resort. However, the real power of the Fourier transform lies in the formal analysis of differential equations.

8.2.2 The Discrete Fourier Transform The preceding remarks alluded to the power of the Fourier transform in solving certain differential equations. If such a solution is combined with numerical techniques, the integrals must be replaced by sums. This is particularly troe if our

218

8. FOURIER ANALYSIS

function is given by a table rather than a mathematical relation, a common feature of numerical analysis. So suppose that we are given a set of measurements performed in equal time intervals of I'>t. Suppose that the overall period in which these measurements are done is T. We are seeking a Fourier transform ofthis finite set of data. First we write

or, discretizing the frequency as well and writing determined later, we have

CUm

= mSco, with I'>cu to be

I N-l T j(ml'>cu) = - f(nl'>t)e- i (mh.w j n 6.t ( - ) •

L

v'2IT n~O

(8.33)

N

Since the Fourier transform is given in terms of a finite sum, let us explore the idea of writing the inverse transform also as a sum. So, multiply both sides of the above equation by [ei (m h.w j k 6.t /(v'2IT)]l'>cu and sum over m:

I -L j(ml'>cu)e N-l

v'2IT m~O

i (m h.w)k 6.t I'>cu

T I'> N-l N-l , = ~ f(n!'>t)e im h.w6.t (k- n ) 2:n: N n~O m~O

LL

T I'> N-l N-l = ~ f(nl'>t) eim h.w6.t (k - n ) . 2:n:N n~O m=O

L

L

Problem 8.2 shows that

N

N-l

~

e im !!J..w 8.t (k - n)

= .

j

if k = n,

e i N t1wt:J.t(k-n) _ eili.wAt(k n) _

1

1

if k

#

n.

We want the sum to vanish whenk # n. This suggests demanding that N I'>cul'>t(kn) be an integer multiple of 2:n:. Since I'>cu and I'>t are to be independent of this (arbitrary) integer (as well as k and n), we must write N I'>cul'>t(k - n) = 2:n:(k - n)

discrete Fourier

=}

T N I'>cu N = 2:n:

=}

I'>cu =

2:n:

T

With this choice, we have the following discreteFouriertransforms:

transforms

2:n:j cu·} - T '

(8.34)

8.2 THE FOURIER TRANSFORM

fast Fourier transform

219

where we have redefined the new 1 to be .J2rcN / T times the old j. Discrete Fourier transforms are used extensively in numerical calculation of problems in which ordinary Fourier transforms are used. For instance, if a differential equation lends itself to a solution via the Fourier transform as discussed before, then discrete Fourier transforms will give a procedure for fiuding the solution uumerically. Similarly, the frequeucy analysis of signals is nicely handled by discrete Fourier transforms. It turns out that discrete Fourier analysis is very iuteusive computationally. Its present status as a popular tool iu computational physics is due primarily to a very efficieut method of calculatiou known as the fast Fourier transform. 10 a typical Fourier transform, one has to perform a sum of N terms for every point. Since there are N points to transform, the total computational time will be of order N 2 • 10 the fast Fourier transform, one takes N to be even and divides the sum into two other sums, one over the even termsand one over the odd terms. Then the computation time will be of order 2 x (N /2)2, or half the original calculation. Similarly, if N /2 is even, one can further divide the odd and even sums by two and obtain a computation time of 4 x (N / 4)2, or a quarter ofthe original calculation. 10general, if N = 2k , then by dividing the sums consecutively, we end up with N transforms to be performed after k steps. So, the computation time will be kN = N log2 N. For N = 128, the computation time will be 100 log2 128 = 700 as opposed to 1282 R<16, 400, a reduction by a factor of over 20. The fast Fourier transformis indeed fast!

8.2.3

The Fourier Transform of a Distribution

Although one can define the Fourier transform of a distribution in exact analogy to an ordinary function, sometimes it is convenient to define the Fourier transform of the distribution as a linear functional. Let us ignore the distinction between the two variables x and k, and simply define the Fourier transform of a function f : JR ---> JR as

f(u) =

=1

'\! 2rc

1

00 . t f(t)e'U dt.

-00

Now we consider two functions,

(j, g) ==

00 1 00 1

L: -00

=

and g, and note that

100 [= 100 v2n-00 00 1 .]

f(u)g(u)du =

-00

=

f

f(u)

1

g(t) [ -1-

../iii-oo

.

g(t)e-'U'dt] du

-00

f(u)e-'U'du dt

g(t)l(t) dt = (I, g) .

The following definition is motivated by the last equation.

220

8. FOURIER ANALYSIS

er;

8.2.5. Definition. Let


er;

(ifJ, f) = (
= (8, f)- = =

frO)

1

00

1 PC ",21l'

=

f(t) dt

-00

i:(~)f(t)dt=(~,f).

Thus, g = 1/$, as expected. TheFourier transform of }j (x - x') ee 8X ' (x) is given by (8- x" f)

= (8x"

f)

1

-! = f(x) =

1 (_I_ $

PC

1

",,21l'"

00

f(t)e -ix't dt

-00

00

=

e - ' x't ) f(t)dt.

-00

Thus,if ,,(x) = 8(x - x'), then ';;(t)

= (1/ $)e-'x't.

III

8.3 Problems 8.1. Consider the function frO) = L~=-oo 8(0 - Zmn}. (a) Show that f is periodic of period 2". (b) What is the Fourier series expansion for f(O). 8.2. Break:the sum L~~-N ein(O-O') into L;;L; +1 + L~~t. Use the geometric sum formula N r N +l _ 1 ~arn=a~_ _'::'

r-1

L..J

n=O

to obtain

~ e'n(O-O') = e'(O-O') e'N(O-~') -

t:1

e'(O-O) -

1 = ei~(N+l)(O-O') sin[~N(O - Of)]. 1 sin[~(O - 0')]

By changing n to -n or equivalently, (0 - Of) to -(0 - Of) find a similar sum from -N to -I. Now put everythiog together and use the trigonometric identity 2 cos a sin P = sin(a

+ P) -

sin(a -

P)

to show that

~

~ e n~-N

'n(O-O') _

-

sin[(N

+ ~)(O 1

Of)]

sin[Z(O - 0')]

.

8.3 PROBLEMS

221

8.3. Find the Fourier series expansion of the periodic function defined on its fundamental cell as if -1f ~8 < 0, if 0 < 8 ~ tt,

!(8) = {-1(1f+8) 1(1f -8)

8.4. Show that An and B" inEquation (8.2) are real when !(8) is real.

8.5. Find the Fourier series expansion of the periodic function !(8) defined on its fundamental cell, (-1f, zr), as !(8) = cosa8, (a) when a is an integer.

(b) when a is not an integer.

8.6. Find the Fourier series expansion of the periodic function defined on its fundamental cell, (-1f, x), as !(8) = 8. 8.7. Consider the periodic function that is defined on its fundamental cell, (-a, a), as !(x) = [z], (a) Find its Fourier series expansion. (b) Evaluate both sides of the expansion at x = 0, and show that 2 it

=

8~

L.., (2k

k=O

1

+ 1)2 .

(c) Show that the infinite series gives the same result as the function when both are evaluated at x = a. 8.8. Let !(x) = x be a periodic function defined over the interval (0, 2a). Find the Fourier series expansion of f .

8.9. Show that the piecewise parabolic "approximation" to a 2 sin(1fxja) in the interval (-a, a) given by the function !(x) = {4X(a + x) 4x(a-x)

if if

- a ~x ~0 O~x~a

hasthe Fourier series expansion 32a2 ~

!(x) =

---;3 ~

1

(2n

. (2n + 1)1fx

+ 1)3 sm

a

.

Plot !(x), a 2 sin(1fxja), and the series expansion (up to 20 terms) for a = I between -I and +I on the same graph. 8.10. Find the Fourier series expansion of !(8) = 8 2 for 181 < and

00 (_I)" 1f2 ---~­

12 -

L..,

n=l

n2



tt .

Then show that

222

8. FOURIER ANALYSIS

8.11. Find the Fourier series expansion of

f(t) = {Sinwt if 0::; t ::; nf o», o if - rr/w::; t::; O. 8.12. What is the Fourier transform of (a) the constant function f (x) = C, and (b) the Dirac delta function 8(x)? 8.13. Show that (a) if g(x) is real, then g*(k) = g( -k), and (b) if g(x) is even (odd), then g(k) is also even (odd). 8.14. Let gc(x) stand for the single function that is nonzero only on a subinterval of the fundamental cell (a, a + L). Define the function g(x) as 00

g(x) =

L

gc(x - jL).

j=-oo

(a) Show that g(x) is periodic with period L. (b) Find its Fourier transform g(k), and verify that 00

g(k) = Lgc(k)

L

8(kL - 2mrr).

m=-oo

(c) Find the (inverse) transform of g(k), and show that it is the Fourier series of gc(x).

8.15. Evaluate the Fourier transform of if [x] < a, if [x] > a.

g(x) = {b - blxl/a

o

8.16. Let f(6) be a periodic function given by f(6) = I:~-oo ane tlle. Find its Fourier transform /(t). 8.17. Let if if

ItI < T, ItI > T.

Show that /(w) = _1_ {Sin[(W - "'O)T] _ sin[(w

"J2ii

Verify the uncertainty relation !'>w!'>t

+ wa)T]} .

w + "'0

w - wo RJ

4rr.

8.3 PR08LEMS

8.18. if f(x) = g(x

+ a), show that

223

j(k) = e-iakg(k).

8.19. For a > 0 find the Fourier transform of f(x) = e-alxl. Is j(k) symmetric? Is it real? Verify the uncertaioty relations. 8.20. The displacement of a damped harmonic oscillator is given by if t > 0

A e- atei""'t {0

r» =

if t < 0'.

1

Fiod jew) and show that the frequency distribution jj(w) 12 is given by -

A2

2

I

If(w)1 = -2 ( )2 + 2' 7l'ltJ-lVO a convolution theorem

i: i:

i: i:

8.21. Prove the convolution theorem: f(x)g(y -x)dx =

j(k)g(k)e

ikY

dk.

What will this give when y = O? Parseval's relation

8.22. Prove Parseval's relation for Fourier transforms: f(x)g*(x)dx =

j(k)g*(k)dk.

In particular, the norm of a function-with weight function equal to I-is iovariant under Fourier transform,

8.23. Use the completeness relation 1 = La In) (nl and sandwich it between [x) and (x'] to find an expression for the Dirac delta function in terms of an iofinite series of orthonormal functions. 8.24. Use a Fourier transform io three dimensions to find a solution of the Poisson equation: V 2<1>(r) = -4np(r). 8.25. For 9'(x) = 8(x - x'), find q;(y). 8.26. Show that jet) = f( -t). 8.27. The Fourier transform of a distribution 9' is given by

I I: _8 ' (t 00

q;(t) =

n).

n=on!

What is 9'(x)? Hiot: Use ql(x) = 9'( -x) 8.28. For f (x) = LZ~o akxk, show that n

j(u) =

y'2;I: ikak8(k)(u ) , k~O

where

224

8. FOURIER ANALYSIS

Additional Reading 1. Courant, R. and Hilbert, D. Methods of Mathematical Physics, vol. I, Interscience, 1962. The classic book by two masters. This is a very readable book written specifically for physicists. Its treatment of Fourier series and transforms is very clear. 2. DeVries, P. A First Course in Computational Physics, Wiley, 1994. A good discnssion of the fast Fourier transform inclnding some illnstrative computer programs. 3. Reed, M., and Simon', B. Fourier Analysis, Self-Adjointness, Academic Press, 1980. Second volume of a four-volume series. A comprehensive exposition of Fourier analysis with emphasis on operator theory. 4. Richtmyer, R. Principles of Advanced Mathematical Physics, SpringerVerlag, 1978. A two-volume book on mathematical physics written in a formal style, but very useful due to its comprehensiveness and the large number of examples drawn from physics. Chapter 4 discusses Fourier analysis and distributions.

Part III

Complex Analysis

_

j j j j j j j

j

9

_

Complex Calculus

Complex analysis, just like real analysis, deals with questions of continuity, con-

vergence of series,differentiation, integration, andso forth. Thereader is assumed to have been exposed to the algebra of complex numbers.

9.1 ComplexFunctions A complex function is a map f : C --> C, and we write f(z) = w, where both z and w are complex numbers.' The map f can be geometrically thought of as a correspondence between two complex planes, the z-plane and the w-plane. The wplane has a real axis and an imaginary axis, which we can call u and v, respectively. Both u and v are real functions of the coordinates of z, i.e., x and y. Therefore, we may write f(z) = u(x, y)

+ iv(x, y).

(9.1)

This equation gives a unique point (u, v) in the w-plane for each point (x, y) in the z-plane (see Figure 9.1). Under f, regions of the z-plane are mapped onto regions of the w-plane. For instance, a curve in the z-plane may be mapped into a curve in the w-plane. The following example illustrates this point. 9.1.1. Example. Letus investigate thebehavior of a coupleof elementary complexfunctions. In particular, we shall look at the way a line y = mx in the z-plane is mapped into curves in the w-plane. 1Strictly speaking, we shouldwrite f : S -e- C whereS is a subset of the complexplane. The reasonis that most functions arenotdefined fortheentire setof complexnumbers, so that thedomain of suchfunctions is notnecessarily C. Weshallspecify the domain only when it is absolutely necessary. Otherwise, we use the genericnotation f : C -e- C, even though f is defined only on a subsetof C.

228

9. COMPLEX CALCULUS

f z

(u,v)

A map from the z-plane to the w-plane.

Figure 9.1

(a) For

w

(x,y)

w = I(z) = z2, we have

w = (x +iy)2 =x 2 -

i

+2ixy,

x 2 - y2 and v(x, y) = 2xy. For y = mx, i.e., for a line in the z-plane withslope m, theseequations yield u = (1- m 2)x 2 and v = 2mx 2. Eliminating x in these equations, we find v = [2m(1 - m 2)]u. This is a line passing through the origin of the w-plane [see Figore 9.2(a)]. Note that the angle the image line makes with the real axis ofthe w-plane is twice the angle the original line makes with the x-axis. (Shnw this!). (b) The function w = fez) = eZ = ex+iy gives u(x, y) = eX cos y and vex, y) = eX siny. Substituting y = mx, we obtain u = eX cosmx and v = ex sinmx. Unlike part (a), we cannot eliminate x to find v as an explicit function of u, Nevertheless, the last pair of equations areparametric equations of a curve, which we can plot in a uv-plane as shown inFigore9.2(b). III with u(x, y)

=

Limits of complex functions are defined in terms of absolute values. Thus, lim z-+ a f(z) = wo means that given any real number f > 0, we can find a corresponding real number 8 > Osuchthatlf(z)-wol < s whenever Izc-u] < 8. Similarly, we say that a function f is cnntinunus at z = a iflirnz-+a f(z) = f(a).

9.2 Analytic Functions The derivative of a complex function is defined as usual:

9.2.1. Definitinn. Let f : iC --> iC be a complex function. The derivative of f at

zo is df

dz

I Zo

= lim f(zo 6.z-+0

+ L>z) -

f(zo) ,

.6.z

provided that the limit exists and is independent of L>z.

9.2 ANALYTIC FUNCTIONS

229

y z2

a x

2a

-+

u

(aJ

y

eZ

a x

-+ (bJ

Figure 9.2 (a)Themap z2 takes aline withslope anglea and maps it to aline withtwice the angle in the w-plane. (b) The map eZ takes the same line and mapsit to a spiral in the w-plane.

In this definition "independent of /),.z" means independent of /),.X and /),.y (the components of Az) and, therefore, independent of the direction of approach to zoo The restrictions of this definition apply to the real case as well. For instance, the derivative of f(x) = [x] at x = 0 does not exisr' because it approaches + I from the right and - I from the left. It can easily be shown that all the formal rules of differentiation that apply to the real case also apply to the complex case. For example, if f and g are differentiable, then f ± g, f g, and-as long as g is not zero- f / g are also differentiable, and their derivatives are given by the usual rules of differentiation. Example Illustrating path dependence of derivative

9.2.2, Example. Let us examine the derivative of fez) df

I

lim

=

f(1

6.z~O

dz z=l+i

.

hm

=

.8.x--*O .6.y~O

+ i + Ll.z) -

f(t

= x 2 + 2ii at z = I + i:

+ i)

..6.z (I

+ /),.x)2 + 2i(t + Ll.y)2 Ax

t - 2i

+ if:!..y

lim 2Ll.x+4iLl.y+(~x)2+2i(Ll.y)2. Ax + l.6..y

=

8x~O .6.y~O

Let us approach z limit yields df

I

dz z=l+i

=

= I + i along the line y lim 2/),.x + 4imLl.x

.6.x~O

- I

=

m(x - I). Then Ll.y

+ (/),.x)2 + 2im 2(Ll.x)2

Ax

+ im Sx

= m Sx, and the

2+4im 1 +im

20ne canrephrase thisandsaythatthederivative exists. butnot in terms of ordinary functions, rather, in terms of generalized functions-in this case 8(x)---discussed in Chapter6.

230

9. COMPLEX CALCULUS

It follows that we get infinitely manyvaluesfor the derivative depending on thevalue we

assign to m, i.e., depending on the direction along which we approach 1 + i, Thus, the derivative does not exist at z = 1 + i, III

It is clear from the definition that differentiability pnts a severe restriction on fez) because it requires the limit to be the saroe for all paths going through zoo Furthermore, differentiability is a local property: To test whether or not a function fez) is differentiable at zo. we move away from zo by a small aroount Az and check the existence of the limit in Definition 9.2.1. What are the conditions under which a complex function is differentiable? For fez) = u(x, y) + iv(x, y), Definition 9.2.1 yields

I

df = lim {u(xo + AX, YO + Ay) - u(xo, YO) dz zo 8x-+O Ax+iAy dy~O

. v(xo + AX, YO + Ay) - v(xo, YO)} +1 . Ax + iAy If this limit is to exist for all paths, it must exist for the two particular paths on which Ay = 0 (parallel to the x-axis) and Ax = 0 (parallel to the y-axis). For the first path we get

I

df = lim u(xo + AX, Yo) - u(xo, Yo) dz zo llx-+O ~x . Ii v(xo + AX, YO) - v(xo, YO) au +1 m = L\x-+O

!:i.x

I

. I

ax (xo,Yo)

av +1 -

ax (xo,yo)

.

For the second path (Ax = 0), we obtain

I

df = lim u(xo, Yo dz eo 8y-+O

+ Ay) -

u(xo, YO)

i Ay

I

. I' v(xo, Yo + Ay) - v(xo, YO) . au +1 rm =-1 8y-+O iAy ay (xo,Yo)

av I +-

ay (XO,Yo) .

If f is to be differentiable at zo, the derivatives along the two paths must be equal. Equating the real and imaginary parts of both sides of this equation and ignoriog the subscript zo (xo, Yo, or zo is arbitrary), we obtain

au av ax = ay Cauchy-Riemann conditions

and

(9.2)

These two conditions, which are necessary for the differentiability of f, are called the Cauchy-Riemann conditions. An alternative way of writing the Cauchy-Riemann (C-R) conditions is obtained by making the substitution' x = !(z + z*) and y = (z - z") in u(x, y)

3Weuse z" to indicate the complexconjugate of z. Occasionally we mayuse z.

11

9.2 ANALYTIC FUNCTIONS

231

and vex, y), using the chain rule to write Equation (9.2) in terms of z and z", substituting the results in

!L = ~ + i ~ and showing that Equation (9.2) is az* az* az*

equivalent to the single equation af/az* = O. This equation says that

9.2.3. Box. Iff is to be differentiable, it must be independent of z".

Expression for the derivative of a differentiable complex function

If the derivative of f exists, the arguments leading to Equation (9.2) imply that the derivative can be expressed as df au' av av au - = -+i-= --i-. dz ax ax ay ay

(9.3)

The C-R conditions assure us that these two equations are equivalent. The following example illustrates the differentiability of complex functions. 9.2.4. Example. Let us determine whether or not the following functions are differentiable: (a) We have already established that fez) = x 2 + 2iy2 is not differentiable at z = 1+ t. We can now show that it is has no derivative at any point in the complex plane (except at the origin). This is easily seen by noting that U = x 2 and v = 2 y 2 , and that au/ax = 2x of. av/ay = 4y, and the first Cauchy-Riemann condition is not satisfied. The secondC-R condition is satisfied, butthat is notenough. We can also write fez) in terms of z and z": fez)

= G(Z +z*)t +2i [;i (z -

z*)t

= 1(1- 2i)(z2 +z*2) + !(l +2i)zz*.

f (z) has anexplicitdependence on z". Therefore, it is not differentiable. (b) Now consider fez) x 2 - y2 + 2ixy, for which u = x 2 and v

= i = 2xy. The C-R conditions become au/ax = 2x = av/ay and au/ay = -2y = -av/ax. Thus, fez) may be differentiable. Recallthat theC-Rconditions areonly necessary conditions; we havenot shown (bot we will, shortly) that they are also sufficient. To check the dependenceof f on z", sobstitntex = (z + z*)/2 and y = (z - z*)/(2i) in u and v to show that f (z) = z2, andthus there is no z" dependence. (c) Let u(x, y) = eX cos y and vex, y) = eX siny. Then au/ax = eX cosy = av/ay and aulay = _ex sin y = -av lax, and the C-Rconditions aresatisfied. Also, fez) = e" cos y + ie" siny

= eX (cosy + i siny) = e"ely =

x iy e +

= eZ,

andthere is no z" dependence. The requirement of differentiability is very restrictive: The derivative must exist along infiuitely many paths. On the other hand, the C-R conditions seem deceptively mild: They are derived for only two paths. Nevertheless, the two paths are, in fact, true representatives of all paths; that is, the C-R conditions are not only necessary, but also sufficient:

232

9. COMPLEX CALCULUS

9.2.5. Theorem. The function f(z) = u(x, y) + iv(x, y) is differentiable in a region of the complex plane if and only if the Cauchy-Riemann conditions,

. au av = ax ay

au = ay

and

av ax

(or, equivalently, af/az* = 0), are satisfied and all first partial derivatives of u and v are continuous in that region. In that case

df dz

au ax

. av ax

av ay

. au ay

-=-+Z-=--z-.

Proof We have already shown the "only if" part. To show the "if" part, note that if the derivative exists at all, it must equal (9.3). Thus, we have to show that

lim f(z

+ M) -

f(z) = au

+ i av

ax

ax

llz

..6..z--+o

or, equivaleutly, that

f (z + A Z) - f (z ) _ (au +iav)1 Az ax ax

I


whenever

IAzl < 8.

By defiuition,

f(z

+ Az) -

f(z) = u(x

+ AX, y + Ay) + iv(x + AX, y + Ay) -

u(x, y) - iv(x, y).

Since u and v havecontinuous first partial derivatives, we can write u(x

+ AX, y + Ay)

v(x

+ AX, y + Ay) =

= u(x, y)

v(x, y)

au

au

av

av

+ ax AX + ay Ay + flAx + 81 Ay, + -Ax + -Ay + f2Ax + 82Ay, ax ay

where sj , ez, 8J, and 82are real numbers that approach zero as Ax and Ay approach

zero.Using these expressions, we can write f(z

+ Az) -

f(z) = (aU ax

+ i av) AX + i (-i au + av) Ay ax

ay

ay

+ (fl + if2)Ax + (81 + i82)Ay au

= ( ax +i

av) (Ax+iAy)+fAx+8Ay, ax

where f == er + iez. 8 == 81 + i82, and we used the C-R conditions in the last step. Dividing both sides by Az = Ax + i Ay, we get

f(z

+ Az) Az

f(z) _ (aU ax

+ i av) = ax

fAX Az

+ 8 Ay. Az

9.2 ANALYTIC FUNCTIONS

233

By the triangle inequality, IRHSI :S lEI + iE21 + 181 + i821. This follows from the fact that Il!.xl/Il!.zl and Il!.yl/Il!.z I are both equal to at most 1. The E and 8 terms can be made as small as desired by making l!.z small enough. We have thus established that when the CoR conditions hold, the function f is differentiable. D

Angustin-Louis Cauchy (1789-1857) was ODe of the most influentialFrench mathematicians of thenineteenth century. He began his career as a military engineer, butwhen his healthbroke down in 1813 he followed his natural inclination and devoted himself wholly to mathematics. In mathematical productivity Cauchy was surpassed only by Euler, andhis collectedworks fill 27 fat volumes. He made substantial contributions to number theory and determinants; is considered to be the originator of the theory of finite groups; and did extensive workin astronomy, mechanics, optics,andthetheory of elasticity.

His greatest achievements, however, lay in the field of analysis. Together with his contemporaries Gauss andAbel, he wasa pioneer in therigorous treatment of limits,continuous functions, derivatives, integrals, andinfinite series.Severalof thebasictestsfortheconvergence of series are associated withhis name. He also provided the first existence proof for solutions of differential equations, gavethefirst proofof theconvergence of aTaylor series, andwas thefirst to feel theneedfor a careful studyof theconvergence behavior of Fourier series (see Chapter 8). However, his most important work was in the theory of functions of a complexvariable, which in essence he created and whichhas continued to be one of the dominant branches of bothpureandapplied mathematics. In this field, Cauchy's integraltheorem andCauchy's integral formula are fundamental tools without whichmodem analysis could hardly exist (see Chapter 9). Unfortunately, his personality did not harmonize with the fruitful powerof his mind. He was an arrogant royalist in politics and a self-righteous, preaching, pious believer in religion-all this in an age of republican skepticism-and most of his fellow scientists disliked him and considered him a smug hypocrite. It might be fairer to put first things first and describe him as a great mathematician who happened also to be a sincere but narrow-minded bigot.

analyticity and singularity; regular and singular points; entire functions

9.2.6. Definition. Afunction f : C --+ C is called analytic at zo if it is differentiable at zo and at all other points in some neighborhood of zoo A point at which f is analytic is called a regular point of f. A point at which f is not analytic is called a singular point or a singularity of f. A function for which all points in iC are regular is called an entire function. 9.2.7. Example.

DERNATIVES OF SOME FUNCTIONS

fez) = z. Hereu = x and v = y; theC-Rconditions areeasily shown to hold, andforany z, we have df/dz = ou/ox +iov/ox = 1. Therefore, the derivativeexists at all pointsof the cornpiex

(a)

234

9. COMPLEX CALCULUS

f(z) = z2

plane. (b)

Here u = x 2 - y2 and v = 2xy; the C-R conditions hold, and for all points z of the complex plane, we have dfldz = aulax + iavlax = 2x + i2y = Zz. Therefore, f(z) is differentiable at all points. (c) f(z) = zn for n 2: I. Wecanusemathematical induction andthefactthat theproduct of twoentire functions is an entire function to show that ..
f(z) = ao + aJZ + ... +an_JZ n-1 +

anZn,

where ai arearbitrary constants. That fez) is entire follows directly from part (e) andthe factthatthesum of two entire functions is entire. (e) fez) = liz. The derivative canbe found to be f'(z) = -Ilz2, which does not exist for z = O. Thus, z = 0 is a singularity of f(z). However, any other point is a regular point of f. (f) f(z) = [z[2. Using the definition of the derivative, we obtain !;.f !;.z

=

[z + !;.z[2 _[z[2 !;,z

(.z:....:,+..:!;.=z)cc(z::.*--,:.+..:!;.=z::.*-')_---..::zz"--* * A * !;.z* =z +uz +z--. !;,z !;.z

For z = 0, !;.fl!;.z = !;.z*, which goes to zero as Az --> O. Therefore, dfldz = 0 at z = 0.4 However, if z i' 0, the limit of !;.fI!;.z will depend on how z is approached. Thus, df/dz does not exist if z i' O. This shows that [z[2 is differentiable only at z = 0 and nowhere else in its neighborhood. It also shows that even if the real (here, u = x2 + ;)

andimaginary (here, v = 0) parts of a complexfunction havecontinuous partial derivatives of alI orders at a point, the functiun may not he differentiable there. (g) f(z) = II sin z: This gives dfldz = - coszl sin 2 z. Thus, f has infinitely many (isolated) singular points at z = ±nrr forn = 0, 1, 2,.... III

9.2.8. Example. THE COMPLEX EXPONENTIAL FUNCTION In this example, we lind the (unique) function f :
(a)

f

is single-valued and analytic for all z,

(b) dfldz (c) f(ZI

= f(z), and

+ Z2) =

f(ZI)f(Z2)·

Property (b) shows that if f (z) is well behaved, then dfI dds also well behaved. In particular, if f (z) is defined for all values of z, then f must he entire. For zt = 0 = Z2, property (c) yields flO) = [/(0)]2 =} f(O) = I, or f(O) = o. On the other hand,

df dz

=

lim

f(z

..6.z~O

+ !;.z) -

f(z)

.6.z

=

lim ..6.z--+0

f(z)f(!;.z) - f(z) 8z

=

f(z) lim f(!;.z) - I ..6.z-+0 ilz

Properly (b) now implies that

lim ..6.z--+0

f(!;.z) - I = I ..6.z

=}

1'(0) = I

and

f(O) = I.

4Although thederivative of Izl 2 exists at z = 0, itis not analytic there(oranywhere else). Tobe analytic at a point,a function musthavederivatives at all points in some neighborhood of the givenpoint.

9.2 ANALYTIC FUNCTIONS

235

The first implication follows from the definition of derivative, and the second from the fact that the only other choice,namely 1(0) = 0, would yield -00 for the limit. Now,we write I(z) u(x, y) + iv(x, y}, for which property(b) becomes

=

au av au av -+i-=u+iv;;;} -=u, -=v. ax ax ax ax These equatious have the most general solution u(x, y) = a(y)e' and v(x, y) = b(y)eX, where a(y) and bey) are the "constants" of integration. The Cauchy-Riemann conditions

now yield aryl = dbldy and daldy = -b(y), whose most general solution is aryl = Acosy+Bsiny,b(y) = Asiny-Bcosy.Ontheotherhand,/(O) = lyieldsu(O,O) = I and v(O, 0) = 0, implyingthat a(O) = I, b(O) = 0 or A = I, B = O. Wethereforeconclude that I(z) = a(y)e X + ib(y)e' = eX(cos y + i sin y) = exi y = eZ •

ex

and e iy are well-defined in the entire complex plane. Hence, e Z is defined and differentiable over all C; therefore, it is entire. II

Both

Example 9.2.7 shows that any polynomial in z is entire. Example 9.2.8 shows that the exponential function e Z is also entire. Therefore, any product and/or sum of polynomials and eZ will also he entire. We can bnild other entire functions. For instance, e iz and e- iz are entire functions; therefore. the trigonometric functions, defined as eiz _ e-iz eiz + e-iz sinz= 2i and cosz= 2 (9.4) are also entire functions. Problem 9.5 shows that sin z and cos z have only real zeros. The hyperholic functions can be defined similarly: sinh z =

eZ _ e- z 2

and

cosh z

=

eZ + e- z 2

(9.5)

Although the sum and the product of entire functions are entire, the ratio, in general, is not. For instance, if f(z) and g(z) are polynontials of degrees m and n, respectively, then for n > 0, the ratio f (z) I g(z) is not entire, because at the zeros of g(z)-which always exist and we assume that it is not a zero of f(z)-the derivative is not defined. The functions u (x, y) and v(x, y)' of an analytic function have an interesting property that the following example investigates. 9.2.9. Example. The familyof curves u (x, y) = constant is perpendicularto the family of curves vex, y) = constant at each point of the complex plane where fez) = u + iv is analytic. This can easily be seenby looking at the normal to the curves.The normal to the curve u(x, y) = constant is simply Vu = (aulax, aulay). Similarly, the normal to the curve v(x, y) = constant is Vv = (avlax, avlay). Takingthe dot productof thesetwononna1s, we obtain au av au av au (au) au (au) (Vu) . (Vv) = ax ax + ay ay = ax - ay + ay ax = 0 by the C-R conditions.

III

236

9. COMPLEX CALCULUS

9.3 Conformal Maps The real and imaginary parts of an analytic function separately satisfy the twodimensional Laplace's equation:

a2 u a2u ax ay 2 - ,

a2 v a2 v ax ai

-+ --0 2

-+ -=0 2

(9.6)

This can easily be verified from the C-R conditions. Laplace's equation in three dimensions,

a2 <1> ax

a2 <1> ai

a2 <1> 8z

-+ -+-0 2 2

harmonic functions

complex potential

-

,

describes the electrostatic potential in a charge-free region of space. In a typical electrostatic problem the potential is given at certain bonndaries (usually conducting surfaces), and its value at every point in space is sought. There are numerous techniques for solving such problems, and some of them will be discussed later in the book. However, some of these problems have a certain degree of symmetry that reduces them to two-dimensional problems. In such cases, the theory of analytic functions can be extremely helpful. The symmetry mentioned above is cylindrical symmetry, where the potential is known a priori to be independent of the z-coordinate (the axis of symmetry). This situation occurs when conductors are cylinders and-if there are charge distributions in certain regions of space-the densities are z-independent. In such cases, ajaz = 0, and the problem reduces to a two-dimensional one. Functions satisfying Laplace's equation are called harmonic functions. Thus, the electrostatic potential is a three-dimensioual harmonic function, and the potential for a cylindrically symmetric charge distribution and boundary condition is a two-dimensional harmonic function. Since the real and the imaginary parts of a complex analytic function are also harmonic, techniques of complex analysis are sometimes useful in solving electrostatic problems with cylindrical symmetry.' To illustrate the connection between electrostatics and complex analysis, consider a long straight filament with a constant linear charge density A. It is shown in introdnctory electromagnetism that the potential (disregarding the arbitrary constant that determines the reference potential) is given, in cylindrical coordinates, by = 2Alnp = 2AIn[(X2 + y2)1/2] = 2AIn [z]. Since satisfies Laplace's equation, we conclude that could be the real part of an analytic function w(z), which we call the complex potential. Example 9.2.9, plus the fact that the curves u = = constant are circles, imply that the constant-v curves are rays, i.e., v ()( rp. Choosing the constant of proportionality as 2A, we obtain w(z) = 2Alnp + i2Arp = 2AIn(pei~). = 2AlnZ

5We use electrostatics because it is morefamiliar tophysicsstudents. Engineering students arefamiliar withsteadystate heat transfer as well, whichalso involves Laplace's equation, andtherefore is amenable to this technique.

9.3 CONFORMAL MAPS

237

It is useful to know the complex potential of more than one filament of charge. To find such a potential we must first find w(z) for aline charge when it is displaced from the origin. If the line is located at zo = xo + iyo, then it is easy to show that w(z) = 2Aln(Z - zo). If there aren line charges located at zjv zj , ... , Zn, then n

w(z) = 2

I>k In(z - Zk)·

(9.7)

k=l

conformal mapping

The function w(z) can be used directly to solve a number of electrostatic problems involving simple charge distributions and conductor arrangements. Some of these are illustrated in problems at the end of this chapter. Iostead of treating w(z) as a complex potential, let us look at it as a map from the z-plane (or xy-plane) to the w-plane (or uv-plane). In particular, the equipotential curves (circles) are mapped onto lines parallel to the v-axis in the w-plane. This is so because equipotential curves are defined by u = constant. Similarly, the constant-v curves are mapped onto horizontal lines in the w-plane. This is an enormous simplification of the geometry. Straight lines, especially when they are parallel to axes, are by far simpler geometrical objects than circles," especially if the circles are not centered at the origin. So let us consider two complex "worlds." One is represented by the xy-plane and denoted by z. The other, the "prime world," is represented? by z', and its real and imaginary parts by x' and y'. We start in z, where we need to find a physical quantity such as the electrostatic potential (x, y). If the problem is too complicated in the z-world, we transfer it to the z'-world, in which it may be easily solvable; we solve the problem there (in terms of x' and y') and then transfer back to the z-world (x and y). The mapping that relates z and z' must be cleverly chosen. Otherwise, there is no guarantee that the problem will simplify. Two conditions are necessary for the above strategy to work. First, the differential equation describing the physics must not get more complicated with the transfer to z'. Since Laplace's equation is already of the simplest type, the z'-world must also respect Laplace's equation. Second, and more importantly, the mapping must preserve the angles between curves. This is necessary because we want the equipotential curves and the field lines to be perpendicular in both worlds. A mapping that preserves the angle between two curves at a given point is called a conformal mapping. We already have such mappings at our disposal, as the following proposition shows. 9.3.1. Proposition. Let VI and Y2 be curves in the complex z-plane that intersectat apoint ui at an angle a.Let f : iC --+ C be a mapping given by f(z) = z' = x'+iy' that is analytic at zoo Let and V~ be the images of VI and Y2 under this mapping, which intersect at an angle a', Then,

vi

6This statement is valid only in Cartesian coordinates. But these are precisely the coordinates we are using in this discussion. 7We are using Zl instead of w, and (x', y') instead of Cu, v).

238

9. COMPLEX CALCULUS

(a) a' = a, that is, the mapping f is conformal,

if (dz'ldzlz o oF O.

(b) Iff is harmonic in (x, y), it is also harmonic in (x', y'). Proof We sketch the proof of the first part. The details, as well as the proof of the second part, involve partial differentiation and the chain rule and are left for the reader. The angle between the two curves is obtained by taking the inner product of the two unit vectors tangent to the curves at zoo A small displacement along Yi can be written as exAxi + eyD..Yi for i = 1,2, and the unitvectors as

'"

ej

=

ex .6.xi

+ eylJ.Yi + (!;.Yi)2

for i = 1,2.

J(!;'Xi)2 Therefore,

, , el . ez =

.

J (!;.XIP

!;.XI!;.X2

+ !;.YI !;.Y2

+ (!;'YI)2J (!;.X2)2 + (!;.Y2P

.

Sintilarly, in the prime plane, we have

e~

. e; =

!;.x; !;.x~ + !;.Y;!;.y~ ,,===='==~r=~==~=== J(!;.x;)2 + (!;.y;pJ(!;.x~)2 + (!;.y~)2'

where x' = u(x, y) andy' = vex, y) andu and v are the real andimaginaryPj of the analytic function f. Using the relations

, !;.x,

au au = -!;.x, + -a !;.Yi, ax Y

, !;.Yi

av av = -a !;.Xi + -!;.Yi, x ay

i

= 1,2,

and the Cauchy-Riemann conditions, the reader may verify that 1'; .e~ = 1'1 ·1'2.

D

The following are some examples of conformal mappings. (a) translation dilation

inversion

z' = z + a, where a is an arbitrary complex constant. This is simply a trans-

lation of the z-plane, (b) z' = bz; where b is an arbitrary complex constant. This is a dilation whereby distances are dilated by a factor Ibl. A graph in the z-plane is mapped onto a similar (congruent) graph in the z'·plane that will be reduced (Ibl < 1) or enlarged

(!bl > 1) by a factor of Ibl. (c) z' = liz. This is called an inversion. Example 9.3.2 will show that under such a mapping, circles are mapped onto circles or straight lines. (d) Combining the preceding three transformations yields the general mapping ,

az+b cz+d

Z=--,

which is conformal if cz + d toad - be oF O.

(9.8) .

oF 0 oF dz'ldz. The latter conditions are eqnivalent

9.3 CONFORMAL MAPS

239

9.3.2. Example, A circle of radius r whose center is at a in the a-plane is described by the equation lz - a I = r. When transforming to the z' -plane under inversion, this equation becomes II/z' - a] = r, or 11- all = rill. Squaring both sides and simplifying yields (r 2 - lal2)lz'1 2 + 2 Re(al) - I O. In terms of Cartesiancoordinates, this becomes

=

(9.9) where a

1. r

== a- + ia;. We now consider two cases:

i' 14 Divideby r 2 -lal 2 and comptetethe squaresto get

' ar )2 (, (x + r 2 _ lal 2 + y -

ai

r 2 - lal2

)2-

a;+af (r2 _ la12)2

-

1

-0 r 2 - lal2 - ,

or definiug a; sa -ar/(r 2 -laI 2 ) , a; ea ai!(r 2 -laI 2 ) , and r' '" r/lr 2 -lal 2l, we have (x' - a;)2 + (y' - aj)2 = r ll, which can also be written as

Ii - a'l = r',

I

,

a = ar

.,

a*

+ la; = lal2 - r2'

This is a circle in the z' -plane with center at a' and radius of r',

2. r = a: Then Equation (9.9) reduces to a-x' - aj y' =

l. which is the equation of a

liue. If we use the transformation z' = 1/(z - c) instead of z' = lIz, then lz - at = r becomes 11/z' - (a - c) I = r, and all the above analysis will go through exactly as before,

except that a is replacedby a - c.

homographic transformations

III

Mappiugs of the form given in Equation (9.8) are called homographic transformations. A nseful property of such transformations is that they can map an iufinite region of the z-plane onto a finite region of the z'-plane. In fact, points with very large values of z are mapped onto a neighborhood of the point z' = a/ c. Of course, this argmnent goes both ways: Eqnation (9.8) also maps a neighborhood of -d/ c in the z-plane onto large regions of the z'-plane. The nsefulness of homographic transformations is illustrated in the followiug example. 9.3.3. Example. Cousidertwo cyliudricalconductorsof equal rar1ius r, beld at potentials Ul and U2. respectively, whose centers are D units of length apart. Choose the x-and the y-axes such that the centers of the cylinders are located on the x-axis at distances al and

«a from the origin, as shown in Figure 9.3. Let us find the electrostatic potential produced

by such a configuration iu the xy-plane. We know from elementary electrostatics that the problem becomes very simple if the two cylinders are concentric (and, of course, of different radii). Thus, we try to map the two circles onto two concentric circles in the z' -plane such that the infinite region outside the two circles in the z-plane gets mapped onto the finite annular region between the two concentric circles in the zl-plane. We then (easily) find the potential in the z'-plane, and transfer it back to the z-plane. The most general mapping that may be able to do the job is that given by Equation (9.8). However, it turns out that we do not have to be this general. In fact, the special case

240

9. COMPLEX CALCULUS

Figure 9.3

In the z-plane, we see two equalcylinders whose centers areseparated.

z' = I/(z - c) in which c is a real constant will be snfficient. So, z = (I/z') + c, and the 1,2 will be mapped onto the circles [z' - a,(1 1"'(, where (by circles Iz -akl I" fork Example 9.3.2) a,( = (ak - c)/[(ak - c)2 - r 2] and r,( = r/I(ak - c)2 _ r 2 1. Can we arrange the parameters so thatthe circles in the z' -plane are concentric, i.e., thata~ a ? Theanswer is yes. We setal = a2 andsolve foraz in terms of al' Theresult

=

=

=

= 2

is either the trivial solution az = at, or cz = c - r 2/(al - c). IT we place the originof the z-plane at the center of the first cylinder, thenat = 0 and az = D = c + r 2 [c. Wecan also findat and a2:at = a2ee a' = -c/(c 2 - r 2 ) , andthe geometry of the problem is as shown in Figure 9.4. For such a geometry the potential at a point in the annular region is given by 4>' = A In p + B = A In Iz' - a'i + B, where A and B are real constants determined by the conditions <1>' (r;) = "1 and <1>' (r~) = "2, which yields

and

Thepotential $' is therealpart of the complex functtonf

F(Z')

= Aln(z' -

a')

+ B,

whichis analytic exceptat z' = a', a pointlying outside theregionof interest. Wecannow go back to the z-plane by substituting z' = I/(z - c) to obtain G(z) = Ain (_1__

z-c

8Writing Z value.

= Izle iO, we note thatln z = In [z] + ie,

a') + B, so that therealpart of a complexlog function is thelog of the absolute

9.4 INTEGRATION OF COMPLEX FUNCTIONS

Figure 9.4

241

"

In ~~ z' -plane, we see two concentric unequal cylinders.

whose realpart is the potential in the z-plane: (x, y)

a'z + a'c l + B = Re[G(z)] = Aln I - z-c I = Alnl(1 +a'c-a'x) -ia'yl +B A =-In 2

[(I

(x-c)+iy +a'c -a'x)2 +a'2y2] (x - c)

2

+ y2

+B. iii

This is the potential we want.

9.4 Integration of Complex Functions The derivative of a complex function is an important concept and, as the previous section demonstrated, provides a powerful tool in physical applications. The concept of integration is even more important. In fact, we will see in the next section that derivatives can be written in terms of integrals. We will study integrals of complex functions in detail in this section. The definite integral of a complex function is defined in analogy to that of a real function: '2

1 at

N

!(z) dz =

L !(Zi)Ll.Zi, N....:,-oo. lim

..6.zr+ O ' =

1

242

9. COMPLEX CALCULUS

where I1Zi is a small segment, situated at zs,ofthe curve that conoects the complex number al to the complex number a2 in the z-plane, Since there are infinitely many ways of connecting al to a2, it is possible to obtain different values for the integral for different paths.

One encounters a similar situation when one triesto evaluate the line integral of a vector field. In fact, we can tum the integral of a complex function into a line integral as follows. We substitute f(z) = u +iv anddz = dx +idy in the integral to obtain

c

f(z)dz =

at

c

(udx - vdy)

at

+i

c

(vdx

+ udy).

at

If we define the two-dimensional vectors Al sa (u, -v) andA2 '" (v, u), then we ra, Al . dr + ira, A2 . dr. It follows from Stokes' theorem get JrUa,I f(z) dz = JUI jUt (or Green's theorem, since the vectors lie in a plane) that the integral of f is pathindependent only if both A I and A2 have vanishing curls. This in tum follows if and only if u and v satisfy the CoR conditions, and this is exactly what is needed for f (z) to be analytic. Path-independence of a line integral of a vector A is equivalent to the vanishing of the integral along a closed path, and the latter is equivalent to the vanishing of V x A = 0 at every point of the region bordered by the closed path. The preceding discussion is encapsnlated in an important theorem, which we shall state shortly. First, however, it is worthwhile to become familiar with some terminology used frequently in complex analysis. curve defined

1. A curve is a map Y : [a, b] -> iC from the real interval into the complex plane given by y(t) = Yr(t) +iYi(t), where a :s: t :s: b, and Yr and Yi are the real and imaginary parts of Y; Y (a) is called the initial point of the curve and y(b) its final point.

2. A simple arc, or a Jordan arc, is a curvethat does not cross itself, i.e., y is injective (or one to one), so that y(tl) "" y(t2) when tl "" ti. 3. A path is a finite collection In, n, ... , Yn} of simple arcs such that the initial point of JIk+1 coincides with the final point of JIk. 4. A smooth arc is a curve for which dy ldt = dYr/dt is nonzero for t E [a, b].

+ idYi/dt exists and

contour defined

5. A contour is a path whose arcs are smooth. When the initial point of YI coincides with the final point of Yn, the contour is said to be a simple closed contour.

Cauchy-Goursat theorem

9.4,1. Theorem. (Cauchy-Goursat theorem) Let f : C -> iC be analytic on a simple closed contour C and at all points inside C. Then

i

f(z)dz = O.

9.4 INTEGRATION OF COMPLEX FUNCTIONS

243

(1,2)

y

x Figure 9.5

Thethree different paths of integration corresponding to theintegrals II. If.

Iz. and I~. 9.4.2. Example.

EXAMPLES OF DEFINITE INTEGRALS

(a) Let us evaluate the integral II = fY1 z dz where ·YI is the straight line drawn from the origin to the point (1.2) (see Figure 9.5). Along such a line y = 2x and. using t for x. YI(t) = t +2it where 0 :s t:s 1; so II

=

1

z dz

)'1

=

[\t

io

+ 2it)(dt + 2idt) =

[I (-3tdt

10

+ 4itdt) = -~ + u.

For a different path)'2. along which y = 2x 2• we get )'2(t) = t and

Ii =

Ir,

zdz

= 10

1

(t

+ 2it 2 where 0 :s t :s

1.

+ 2it 2)(dt + 4itdt) = -~ + 2i.

Therefore, II = If. This is what is expected from the Cauchy-Goursat theorem because the function fez) = z is analytic on the two paths and in theregionbounded by them. (h) To find Iz es f z 2dz with YI as in part (a). suhstitute for z in terms of t: Y1

Iz =

1 2dt

1n (t + 2it)2(dt + 2idt) = (1 + 2i)3 10 t

11

2 -3 i,

= --3 -

0

Next we compare h with I~ = f Y3 be described by

n(t) =

g

for 0

+i(t -1)

z 2d z where Y3 is as shown in Figure 9.5. This pathcan

s

t .::: 1,

for 1 s t ::::: 3.

Therefore,

,= 10o1 t2dt + /,31

12

[1 + itt - 1)]2 (idt)

= -1 3

2 4 - -i 3

2 = - 11 - - -zt,

3

3

244

9. COMPLEX CALCULUS

Figure 9.6

The two semicircular paths for calculating h and I~.

which is identical to 12.once againbecausethefunction is analytic on Yl and Y3 as well as in theregionbounded by them. (c) Now consider /3 es fY4 dz/z where Y4 is the upper semicircle of unit radius, as shown in Figure 9.6. A parametric equation for Y4 can be given in terms of (): O:::::8::S1L Thus,we obtain

On the other hand,

,1

13 =

t = -dz

Y~ Z

j,"

t . tedO =

~,e

2n el

.

-llr.

Herethetwo integrals arenot equal. From Y4 and Y~ we can construct a counterclockwise simpte closed contour C, along which the integral of f(z) = l/z becomes Jc dz/z = /3 - I~ = liT(. Thatthe integral is not zero is a consequence of the fact that liz is not analytic at all pointsof theregionbounded by the closed contour C. ..

The Cauchy-Goursat theorem applies to more complicated regions. When a region contains points at which f(z) is not analytic, those points can be avoided by redefining the region and the contour. Such a procedure requires an agreement on the direction we will take. convention for positive sense of integretion around a closed contour

Convention. When integrating along a closed contour, we agree to move along the contour in such a way that the enclosed region lies to our left. An integration thatfollows this convention is called integration in the positive sense. Integration performed in the opposite direction acquires a minus sign.

9.4 INTEGRATION OF COMPLEX FUNCTIONS

245

y

o

x

Figure 9.7 A complicated contour can be broken up into simpler ones. Note that the boundaries of the "eyes" and the "mouth" are forced to be traversed in the (negative) clockwise direction.

For a simple closed contour, movement in the counterclockwise direction yields integration in the positive sense. However, as the contour becomes more compli-

cated, this conclusion breaks down. Figure 9.7 shows a complicated path enclosing a region (shaded) in which the integrand is analytic. Note that it is possible to traverse a portion of the region twice in opposite directions without affecting the integral, which may be a sum of integrals for different pieces of the contour. Also

simply and multiply connected regions

Cauchy Integral Formula

note that the "eyes" and the "mouth" are traversed clockwise! This is necessary because of the convention above. A region such as that shown in Figure 9.7, in which holes are "punched out," is called multiply connected. In contrast, a simply connected region is one in which every simple closed contour encloses only points of the region. One important consequence of the Cauchy-Goursat theorem is the following: 9.4.3. Theorem. (Cauchy iutegral formula) Let ! be analytic on and within a simple closed contour C integrated in the positive sense. Let zo be any interior point to C. Then I !(ZO) = - .

i

!(z) --dz 2:m c z - zo

To prove the Cauchy integral formula (CIF), we need the following lemma. 9.4.4. Lemma. (Darboux inequality) Suppose ! : C -+ iC is continuous and bounded on a path y, i.e.•there exists a positive number M such that 1!(z)1 ::: M

246

9. COMPLEX CALCULUS

for all values z E y. Then

where L y is the length ofthe path of integration. Proof

11

f(Z)dZI

y

= IN~ootf(Zi)!!.Zil = J!!!'oo Li.Zi--+O l=l

s

It I f(Zi)!!.Zi

.6.Zj--+O 1=1

N lim ''If(zi)!!.ziI =

!--

N--+oo ..6.zi--+O 1=1

N

L If(Zi)1

lim

N--+oo . 8Zj--+O 1=1

l!!.ziI

N

sM

lim

L l!!.ziI = ML y •

N-*oo. 1

Li.zi--+O 1=

The first inequality follows from the triangle inequality, the second from the boundedness of f, and the last equality follows .from the definition of the length of a path. D Now we are ready to prove the Cauchy integral formula.

ProofofCIF. Consider the shaded region in Figure 9.8, which is bounded by C, by YO (a circle of arbitrarily small radius 8 centered at zo), and by Ll and L2, two straight line segments infinitesimally close to one another (we can, in fact, assnme that L 1 and L2 are right on top of one another; however, they are separated in the figure for clarity). Let us use C' to denote the union of all these curves. Since f(z)/(z - zo) is analytic everywhere on the contour C' and inside the shaded region, we can write

1

0= _1_ f(z) dz 2ni fe! z - zo =

_1_[1 2ni

f(z) dz

rc z - zo

(9.10)

+1

f(z) dz

fro Z -

zo

+1

f(z) dz

hI Z -

+1

f(z) dzJ.

hl Z -

zo

zo ,

=0

The contributions from L 1 and L2 cancel because they are integrals along the sarne line segment in opposite directions. Let us evaluate the contribution from the infinitesimal circle yo.First we note that because f (z) is continuous (differentiability implies continnity), we can write

I

f(z) - f(zo) z-zo

I = If(z) -

f(zo) I

Iz-zol

= If(z) -

f(zo)!

8

•8

<-

9.4 INTEGRATION OF COMPLEX FUNCTIONS

247

y

x Figure 9.8 The integrand is analytic withinandon theboundary of the shaded region. It is alwayspossibleto construct contours that excludeall singular points.

for Z E JIO, where. is a small positive number. We now apply the Darbonx inequality and write

1

1. fro

!(z) - !(zo) dzi < =-2,,8 = 2" •. z-zo 8

This means that the integral goes to zero as 8 ---> 0, or

.ss.;

1.

!(z) dz = 1. !(zo) dz = !(zo) 1. froz-zo froz-zo froz-zo

We can easily calculate the integral on the RHS by noting that z - zo = 8ei • and that JIO has a clockwise direction:

i

Yo

dz

-- = z - zo

1

2 "

0

i·d
i8e . ' qJ 8e

= -2"i

=}

i

Yo

!(z) --dz = -2"i!(zo). Z - zo D

Snbstitnting this in (9.10) yields the desired result. 9.4.5. Example. Wecan use the elf to evaluate the integrals (z2 - I)dz

Iz= fcc, ·(z -

iHz 2 -

4)3

,

248

9. COMPLEX CALCULUS

where CI, e2, andC3 arecircles centered at the origin withradii 71 = 3/2,72 = I, and =4. For lj we note that /(z) = z2/(z2 + 3)2 is analyticwithin and on C\o and zc = i lies in theinterior of C1. Thus, 73

II

=

t

/(z)dz

..

.

i2

.st

- - . = 2",/(,) = 21"'2 2 = -'-2' c, z -, (, + 3)

Similarly, /(z) = (z2 - 1)/(z2 - 4)3 for the integral Iz is analytic on and within C2, and = 1/2 is an interior point of C2. Thus, the Cllt gives

ZQ

Iz =

i

/(z)dz

--1

C,z-:;:

32".

= 2".i/(1/2) = --5 i . 112

For the last integral, / (z) = ez/2/ (z2 - 20)4, and the interiorpoint is zu = iit : 2".

III

The Cauchy integral formula gives the value of an analytic function at every point inside a simple closed contour when it is given the value of the function only at points on the contour. It seems as though an analytic function is not free to

change insidea region once its valueis fixedon the contour enclosingthat region. Explanation at why the Cauchy integral formula works!

There is an analogons sitnation in electrostatics: The specification ofthe potential at the boundaries, such as the surfaces of conductors, automatically determines the potential at any other point in the region of space bounded by the conductors. This is the content of the uniqueness theorem used in electrostatic boundary value problems. However, the electrostatic potential is bound by another condition, Laplace's equation; and the combination of Laplace's equation and the boundary conditions furnishes the uniqueness of <1>. Similarly, the real and imaginary parts of an analytic function separately satisfy Laplace's equation in two dimensions! Thus, it should come as no surprise that the value of an analytic function on a boundary (contour) determines the function at all points inside the boundary.

9.5 Derivatives as Integrals The Cauchy Integral Formula is a very powerful tool for working with analytic functions. One of the applications of this formula is in evaluating the derivatives of such functions. It is convenient to change the dununy integration variable to ~ and write the ClF as f(z) =

---.!..." 1. f(~) d~ , 2"., fc ~ - z

(9.11)

where C is a simple closed contour in the ~ -plane and z is a point within C. As preparation for defining the derivative of an analytic function, we need the following result.

9.5 DERIVATIVES AS INTEGRALS

249

9.5.1. Proposition. Let y be any path---
~ 2:n:z

1 y

g(;) d; ; - z

is analytic at every point z ¢ y. Proof The proof follows immediately from differentiation of the integral: df =_1 ~lg(;)d; =_1 19(;)d;~(_I_)=_11 g(;)d;. dz 2:n:i dz y ; - z 2:n:i y dz ; - z 2:n:i y (; - Z)2 This is defined for all values of z not on y.9 Thus, f(z) is analytic there.

D

We can generalize the formula above to the nth derivative, and obtain n d f nl g(;)d; dz n = 2:n:i y (; - z)n+1.

1

Applying this resultto an analytic function expressed by Equation (9.11), we obtain the following important theorem. derivative of an analytic function given interms of an integral

9.5.2. Theorem. The derivatives of all orders of an analytic function f (z) exist in the domain of analyticity of the function and are themselves analytic in that domain. The nth derivative of f(z) is given by n f(n)(z) = d f = ~ f(;) d; (9.12) dz" 2:n:i ic (; - z)n+1

1

9.5.3. Example. Let us apply Bquation (9.12) directly 10 some simple functions. In all . cases, we will assume that the contour is a circle of radius r centeredat z. (a) Lei f(z) = K, a constant. Then, for n = 1 we have df __1_1 ~ dz - 2"i f c (I; - z)2 . Since ~ is on thecircle C centered at z, g - z

r:

df = _1_ dz 2"i 10 (b) Given

f

riei~ d8. So we have

iO Kire dO = ~ f2rr e-lOso = O. (relO )2 Zatr 10 .

(z) = z, its first derivative will be

df = _1_ 1 ~ = _1_ dz 2ni f c (I; - z)2 2"i =

= re ifJ and dg =

r: (z+relO)irelOdO

10

(re,O)2

~ (: t" e-lOdO + t" -) = ~(O+2") = 2n rio 10 21l'

I.

,

9The interchange of differentiation andintegration requires justification. Such an interchange can be done if the integral has some restrictive properties. We shallnot concern ourselves with such details. In fact, one can achievethe same result by using thedefinition of derivatives andtheusualproperties of integrals.

250

9. COMPLEX CALCULUS

(c) Given f(z) = z2, for the first derivative Equation (9.12) yields df = _1_ 1. g2 dg = _1_ {2n (z + re i9)2ireiOdB dz

'Iitl

=

~

fc (g -

r:

2" 10

z)2

2"i

10

(re,0)2

[z2 + (rei9)2"+ 2zre'0j (reiO)-ldB

I

foh)

=-

(z2 . - fo2n" e-'OdB+r fo2n e'OdB+2z dB ee Zz, r 0 '0 0 It can be shown that, in general, (d/dz)z" = mz..- l. The proof is left as Problem 9.30. III 2n

The CIF is a central formula in complex aualysis, and we shall see its significance in much of the later development of complex analysis. For now, let us demonstrate its usefulness in proving a couple of important properties of analytic functions. 9.5.4. Proposition. The absolute value ofan analytic function f(z) cannot have a local maximum within the region ofanalyticity of the function.

Proof Let S c If(z)1 for all Z on )/0. We now show that this cannot happen. Using the CIF, and noting that z - zo = 8e iO, we have If(zo)1 =

I~,[ 2".,

s~

f(z) dzl = _I

{2n If(z)!de

2" 10

I

{2n f(z)i8e i9de

2" 10

fro z - zo

s

_I

8e'0

r: use

2" 10

I

= M,

where M is the maximum value of If(z)1 for z E )/0. This inequality says that there is at least one point z on the circle )/0 (the point at which the maximum of !f(z)1 is attained) such that If(zo)1 ::: If(z)l. This contradicts our assumption. D Therefore, there can be no local maximum within S. 9.5.5. Proposition. A bounded entire function is necessarily a constant.

Proof We show that the derivative of such a function is zero. Consider df 1,[ f(g) dg dz = 2"i c (g - Z)2 • Since f is an entire function, the closed contour C can be chosen to be a very large circle of radius R with center at z. Taking the absolute value of both sides yields

f

9.5 DERIVATIVES AS INTEGRALS

251

where M is the maximum of the function in the complex plane. Now, as R ---> 00, the derivative goes to zero, and the function must be a constant. D

fundamental theorem ofalgebra proved

Proposition 9.5.5 is a very powerful statement about analytic functions. There are many interesting and nontrivial real functions that are bounded and have derivativesof all orders on theentire real line. Forinstance, e- x2 is sucha function. No such freedom exists for complex analytic functions. Any nontrivial analytic function is either not bounded (goes to infinity somewhere on the complex plane) or not entire (it is not analytic at some point(s) of the complex plane). A consequence of Proposition 9.5.5 is the fundamental theorem of algebra, which states that any polynomial of degree n 2': 1 has n roots (some of which may be repeated). In other words, the polynomial

p(x) = aD +alX + ... +anx n

for n, 2': I

can be factored completely as p(x) = c(x - Zl)(X - zz)··· (x - Zn) where c is a constant and the Zi are, in general. complex numbers. To see how Proposition 9.5.5 implies the fundamental theorem of algebra, we let f(z) = l/p(z) and assume the contrary, i.e., that p(z) is never zero for any (finite) Z E C. Then f(z) is bounded and analytic for all z E C, and Proposition 9.5.5 says that f (z) is a constant. This is obviously wrong. Thus, there mnst be at least one z, say z = zi, for which p(z) is zero. So, we can factor out (z - Zl) from p(z) and write p(z) = (z - Zl)q(Z) where q(z) is of degree n - 1. Applying the above argument to q(z), we have p(z) = (z - Zl)(Z - zz)r(z) where r(z) is of degree n - 2. Continning in this way, we can factor p(z) into linear factors. The last polynomial will be a constant (a polynomial of degree zero) which we have denoted by c. The primitive (indefinite integral) of an analytic function can be defined using definite integrals just as in the real case. Let f : C ---> C be analytic in a region S of the complex plane. Let zo and z be two points in S, and defmel'' F(z) es f(l;) d~. We can show that F(z) is the primitive of f(z) by showing that

fz:

lim

I F(z + t>z) -

.6.z--+O

F(z) - f(z)1 =

o.

f).z

We leave the details as a problem for the reader. 9.5.6. Proposition. Let f : C ---> C be analytic in a region S ofC. Then at every point z E S, there exists an analytic function F : C ---> C such that

dF dz = f(z).

lONate that theintegral is path-independent dueto the analyticity of f. Thus, F is well-defined.

252

9. COMPLEX CALCULUS

In the sketch of the proof of Proposition 9.5.6, we used only the continuity of f and the fact that the integral was well-defined. These two conditions are sufficient to establish the analyticity of F and f, since the latter is the derivative of the former. The following theorem, due to Morera, states this fact and is the converse of the Cauchy-Goursat theorem. 9.5.7. Theorem. (Morera's theorem) Let a function f : iC -+ iC be continuous in a simply connected region S. If for each simple closed contour C in S we have :Fe f(~) d~ = 0, then f is analytic throughout S.

9.6 Taylor and Laurent Series The expansion of functions in terms of polynomials or monomials is important in calculus and was emphasized in the analysis of Chapter 5. We now apply this concept to analytic functions.

9.6.1 Properties of Series

absolute

The reader is assumed to have some familiarity with complex series. Nevertheless, we state (without proof) the most important properties of complex series before discussing Taylor and Laurent series. A complex series is said to converge absolutely if the real series L~o IZkl =

J

convergence

L~o x~ + y~ converges. Clearly, absolute convergence implies convergence.

power series

9.6.1. Proposition. If the power series L~o ak(Z - ZO)k converges for Zl i' zo, then it converges absolutely for every value of Z such that Iz - zol < IZI - zol. Similarly if the power series L~o bk/(Z - ZO)k converges for Z2 i' zo. then it converges absolutely for every value of z such that [z - zol > IZ2 - zn].

circle of convergence

A geometric interpretation of this proposition is that if a power series-with positive powers--converges for a point at a distance 71 from ZQ, then it converges for all interior points of the circle whose center is zo, and whose radius is r\. Similarly, if a power series-with negative powers-converges for a point at a distance r2 from zo, then it converges for all exterior points of the circle whose center is zo and whose radius is r: (see Figure 9.9). Generally speaking, positive powers are used for points inside a circle and negative powers for points outside it. The largest circle about zo such that the first power series of Proposition 9.6.1 converges is called the circle of convergence of the power series. The proposition implies that the series carmot converge at any point outside the circle of convergence. (Why?) In determining the convergence of a power series 00

S(z) es

I>n (z - zo)". n=O

(9.13)

9.6 TAYLOR ANO LAURENT SERIES

253

(a) Figure 9.9 (a) Power series with positive exponents converge for the interior points of a circle. (b) Power series withnegative exponents converge fortheexterior points of a circle.

we look at the behavior of the sequence of partial sums N

SN(Z) es Ean(z - zo)". n=O

Convergence of (9.13) implies that for any 8 > 0, there exists an integer N. such that whenever N > Nsuniform convergence explained

In general, the integer N. may be dependent on z; that is, for different values of z, we may be forced to pick different N. 's. When N. is independent of z, we say that the convergence is uniform.

power series are uniformly convergent and analytic

9,6.2. Theorem. The power series S(z) = L~oan(Z - ao)" is uniformly convergent for all points within its circle of convergence and represents a function that is analytic there. By substituting the reciprocal of (z - zo) in the power series, we can show that if L~o bk/(z - ZO)k is convergent in the annulus rz < Iz - zol < rl, then it is uniformly convergent for all Z in that annulus.

power series can be differentiated and integrated term by term

9.6.3. Theorem. A convergent power series can be differentiated and integrated term by term; that is, if S(z) = L~o an (z - zo)". then dS(z) 00 ~=Enan(z-zo)n-I, n=l

1

00 S(z)dz=Ea n

Y

n=O

1

(z-zo)ndz

Y

for any path y lying in the circle ofconvergence of the power series.

254

9. COMPLEX CALCULUS

Taylor and Laurent Series

9.6.2

We now state and prove the two main theorems of this section. A Taylor series consists of terms with only positive powers. A Laurent series allows for negative powersas well. Taylor series

9.6.4. Theorem. (Taylor series) Let f be analytic throughout the interior of a circle Co having radius ro and centered at zoo Then at each point z inside Co.

f(z) = f(zo)

+ f'(zo)(z -

zo)

L f(n)(zo) (z - zo)". n. 00

+ ... =

I

(9.14)

n=O

That is. the power series converges to f(z) when [z - zol < roo Proof From the CIF and the fact that z is inside Co, we have

~ J f(~) d~. 2rrl feo ~ - z

f(z) =

On the other hand,

1

1

~- z

= =

~

1

- zo + zo - z =

1

(~_ zo)

1

1

(1 _ z --zozo) ~

~

(Z - ZO)n

~-zo 1-~ = ~-zo~ ~-zo ; -zo n_O

The last equality follows from the fact that I(z - zo)/(~ - zo)1 < I-because z is inside the circle Co and ~ is on it-and from the sum of a geometric series. Substituting in the CIF and using Theorem 9.5.2, we obtain the result. D Maclaurin series

For zo = 0 we obtain the Maclanrin series:

f(z)

00

f(n)(o)

n=O

n!

= f(O) + J'(O)z + ... = L __ zn

The Taylor expansion reqnires analyticity of the function at all points interior to the circle Co. On many occasions there may be a point inside Co at which the function is not analytic. The Laurent series accommodates such cases. Laurent series

9.6.5. Theorem. (Laurent series) Let Cl and C2 be circles ofradii Yj and ri. both centered at zo in the Z -plane with Yj > rz. Let f : iC --+ iC be analytic on C1 and C2 and throughout S, the annular region between the two circles. Then, at each point z E S, f(z) is given by 00

f(z) =

L

an(z - zo)"

where

n=-oo

and C is any contour within S that encircles

an = _1 J f(~) d~ 2rri f c (~ - zo)n+l

zo.

9.6 TAYLOR AND LAURENT SERIES

Figure 9.10

255

The annular region within andon whose contour the expanded function is

analytic.

Proof Let Y be a small closed contour in S enclosing z, as shown in Figure 9.10. For the composite contour C' the Cauchy-Goursat theorem gives

0=

J

f(~) d~ =

Yc' ~ - Z

J f(~) d~ _ J f(~) d~ _ J f(~) d~, YCI ~ - Z Yc, ~ - z r, ~ - z

where the y and C2 integrations are negative because their interior lies to our right as we traverse them. The y integral is simply 2rr:if(z) by the CIF. Thus, we obtain

2rr:if(z) =

J f(~) d~ _ J f(~) d~. YCI ~ - Z Yc, ~ - z

(9.15)

Now we use the same trick we used in deriving the Taylor expansion. Since z is located in the annular region, ri < [z - zol < rl. We have to keep this in mind When expanding the fractions. In particular, for ~ E CI we want the ~ term in the denominator, and for ~ E C2 we want it in the numerator. Substituting such expansions in Equation (9.15) yields 00

2rr:if(z) = I:(z - zo)" n=O

i

Cl

+ I: ( I )n+l n=Oz-zo 00

f(~)d~

(~_

i C2

zo

)n+l

(9.16)

f(;)(~ - zo)" d~.

Now we consider an arbitrary contour C in S that encircles zoo Figure 9.11 shows a region bounded by a' contour composed of CI and C. In this region

256

9. COMPLEX CALCULUS

1(1;)/(1; - zo)n+l is analytic (because I; can never equal zo). Thus, the integral over the composite contourmust vanish by the Cauchy-Goursat theorem. It follows that the integral over C 1 is equal to that over C. A similar argument shows that the C2 integral can also be replaced by an integral over C. We let n + 1 = -m iu the second sum of Equation (9.16) to transform it into

f:

m~-1

1

m

(z - zo)

J. 1(1;)(1; 1'c

f

zo)-m-l dl; =

(z - zo)m

m=-oo

J. 1(1;) ~:+1 1'c (I; - zo)

.

Changing the dummy index back to n and substituting the result in Equation (9.16) yields 00

2:n:i/(z) = L(z - zo)"

i

1(1;)

(I; _

C

n=O

zo

)n+l dl;

+

n=-1

L

(z - zo)"

-00

i

c

1(1;)

(I; _

zo

)n+l dl;.

We can now combine the sums and divide both sides by 2:n: i to get the desired expansion. D The Laurent expansion is convergeut as loug as'2 < Iz- Zo I < 'I. Iu particular, if'2 = 0, and if the function is analytic throughout the interior of the larger circle, then an will be zero forn = -1, -2, ... because 1(1;)/(1; _zo)n+l will be analytic for negative n, and the integral will be zero by the Cauchy-Goursat theorem. Thus, ouly positive powers of (z - zo) will be present in the series, and we will recover the Taylor series, as we should. It is clear that we can expand CI and shrink C2 until we encounter a poiut at which I is no longer analytic. This is obvious from the constructiou of the proof, in which ouly the analyticity in the annular region is important, not its size. Thus, we can include all the possible analytic points by expanding CI and shrinking C2. 9.6.6. Example. Letus expand some functions in tenus of series. Foranentire function there is no pointin theentire complex plane at which it is notanalytic. Thus, onlypositive powers of (z - zo) will be present, andwe will havea Taylor expansion that is validfor all values of z. (a) Let us expand e Z around 20 = O. Thenth derivative of e Z is e Z • Thus, j(n)(O) = 1, and Taylor (Maclaurin) expansion gives tIn) (0)

zn

n= L-' e'= n=O L--z n! n=O n! 00

00

(b)TheMaclaurin series for sinz is obtained by noting that d

n

.

I

-r-r-sm z dz n ,~o

= {O(_I)(n-I)/2

ifn is even,

ifnisod
and substituting this in theMaclaurin expansion: n

00

2k+l

(_I)(n-I)/2~ = L(-I)k z n odd n! k~O (2k

sinz = L

+ 1)1

.

9.6 TAYLOR AND LAURENT SERIES

Figure 9.11

257

Thearbitrary contour in the annular regionusedin theLaurent expansion.

Similarly, we canobtain Z2k

00

00

cosz = {;,(_I)k (Zk)!'

Z2k+l

sinhz = (;, ""(Z"-k-+--'I'""')!'

00

coshz =

Z2k

I: (Zk)." k~O

(c) Thefunction 1/ (1 + z) is notentire, so the regionof its convergence is limited. Letus find the Maclaurin expansion of this function. The function is analytic within all circles of radii r < 1. At r = 1 we encounter a singularity, the point z = -1. Thus, the series converges for all pointsl! z forwhich [z] < 1. Forsuchpoints we have n

/(n)(O)

= -d[n( I +z)-I] I dz

z=o

.

= (_I)nnL

Thus, _1_ = f l+z

n=O

/(n) (0) z" = f(-I)nZn. n! n=O

III

Taylor and Laurent series allow us to express an analytic function as a power series. For a Taylor series of !(Z), the expansion is routine because the coefficient of its nth term is simply !(n)(zo)/nl, where zo is the center of the circle of convergence. When a Laurent series is applicable, however, the nth coefficient is not, in general, easy to evaluate. Usually it can be found by inspection and certain manipulations of other known series. But if we use such an intuitive approach 11 As remarked before, the seriesdiverges for all points outside the circle lz! = 1. Thisdoesnotmeanthat thefunction cannot be represented by a seriesfor pointsoutside the circle. On the contrary, we shall see shortly that Laurent series,with negative powers of z - ze aredesignedprecisely for sucha purpose.

258

9. COMPLEX CALCULUS

to determine the coefficients, can we be sure that we have obtained the correct Laurent series? The following theorem answers this question. Laurent series is unique

9.6.7. Theorem. If the series L~-oo an(z - zo)n converges to fez) at all points in someannular region about zo. then it is theunique Laurent series expansion of fez) in that region. Proof Multiply both sides of

f (z) =

L~-oo an(z - zo)n by

1 211:i(z - zo)k+!' integrate the result along a contour C in the annular region, and use the easily verifiable fact that

1 J dz 211:i f c (z - zo)k n+l = 8kn to obtain I J fez) 211:i fc (z _ zo)k+! dz = ar.

Thus, the coefficient in the power series of f is precisely the coefficient in the Laurent series, and the two must be identical. 0 We will look at some examples that illustrate the abstract ideas developed in thepreceding collectionof theorems and propositions. However, we canconsider a much broader range of examples if we know the arithmetic of power series. The following theorem giving arithmetical manipulations with power series is not difficult to prove (see [Chur 74]). You can add, subtract, and multiply convergent power series

9.6.8. Theorem. Letthetwopowerseriesf(z) = L~_ooan(z-zo)nandg(z) = L~-oo bn (z - zo)n be convergent within some annular region r2 < [z- zo I < TJ· Then the sum L~-oo (an+bn)(z - zo)n converges to f (z) + g(z), and the product 00

00

L L n=-oo m=-oo

00

anbm(z - zo)m+n

==

L

Ck(Z - ZO)k

k=-QO

converges to f(z)g(z) for Z interior to the annular region. Furthermore, if g(z) i' 0 for some neighborhood of zo. then the series obtained by long division ofL~_ooan(Z - eo)" by L:'=-oo bm(z - zo)" converges to f(z)/g(z) in that neighborhood. This theorem, in essence, says that converging power series can be manipulated as though they were finite sums (polynomials). Such manipulations are extremely useful when dealing with Taylor .and Laurent expansions in which the straightforward calculation of coefficients may be tedious. The following examples illustrate the power of infinite-series arithmetic.

9.6 TAYLOR AND LAURENT SERIES

9.6.9. Example. Toexpandthefunctionf(z) =

rewrite it as

259

22+3~ in a Laurent series about z = 0,

z +z

= ~21+Z (2+ 3Z) = ~ (3 __1_) = ~ (3 _ ~(_I)nZn)

fez)

Z

z2

I+z

Z2

L.. n=O

1 2 3 2 1 2 =-(3-I+z-z +Z -···)=-+--I+z-z + .... Z2

Z2

. Z

This series converges for 0 < [z] < 1.Wenote that negativepowersof z are also present.12 Using the notation of Theorem 9.6.5, we have a_2 = 2, a_I = 1, an = 0 for n :::s -3, and an = (_1)n+1 forn '" O. III 9.6.10. Example. The function fez) = 1/(4z - z2) is the ratio of two entire functions. Therefore, by Theorem 9.6.8, itis analytic everywhere except at the zeros of its denominator, Z = 0 and z 4. For tile annularregion (here r2 of Theorem 9.6.5 is zero) 0 < lel < 4, we expand f(z) in the Laurent series around z = O. Instead of actually calculating an, we first

=

note that

fez) = :z (I _lz/ 4)· The second factor can be expanded in a geometric series because Iz/41 < 1: OO

n

00

l _ ' " (~) _ ' " 4-n n _ 1 _ z/4 - L.. 4 - L.. z. n=O

n=O

Dividing this by cz, and noting that z = 0 is the only zero of 4z and is exclnded from the annular region, we obtain the expansion

Although we derived this series using manipulations of other series, the uniqueness of series representations assures us that this is the Laurent series for the indicated region. How can we represent fez) in the region for which Iz[ > 4? This region is exterior to the circle lel = 4, so we expect negative powers of z. To find the Laurent expansion we write

fez) = -

z; C_14/J

and note that ]4/zl < 1 for points exterior to the larger circle. The second factor can be written as a geometric series:

12This is a reflection of the fact that the function is not analytic inside the entire circle

lz! =

1; it blows up at z = O.

260

9. COMPLEX CALCULUS

Dividingby _z2, whichis nonzero in the regionexterior to the larger circle,yields 00

f(z) = - I>"z-n-2

11II

n=Q

9.6.11. Example. The function fez) = z([(z -1)(z-2)] has a Taylor expansion around theorigin Iorlz]

<

1. Tofind this expansion, we write13

1

2

1

1

f(z) = - - 1 + - 2 = -1- - - 1 (2· zz-z -z Expanding both fractions in geometric series (hoth [z] and Iz(21 are less than 1), we ohtain fez) = L~o z" - L~O(z(2)n. Adding the two series-nsing Theorem 9.6.8-yie1ds 00

fez) = L(l- 2-")zn

for Izl < 1.

n=O

This is the unique Taylor expansion of fez) within the circle Izi = 1. For 1 < Izi -c 2 we have a Laurent series. To obtain this series, write

l(z 1 fez) = l(z - 1 - 1 - z(2 =

-z

1(

1) 1 1 - l(z - 1 - z(2·

Since hoth fractions on the RHS converge in the annuiarregion (11(zl < 1, Iz(21 < 1), we get 100

fez) =

In

n=D

00

= - Lz-

n=O 00

00

n=O

n=-l

n- I

n=O

L z"> Lz-nzn= L

-00

=-

m n

00

-z L (z) - L

anZ

00

- L2-

nzn

n=O

n,

n=-oo

where an = -1 for n < 0 and an = _Z-n farn ::: O. Thisis the unique Laurent expansion of f (z) in the given region. Finally, for Izl > 2 we have only negative powers of z. We obtain theexpansion in this region hy rewriting f (z) as follows: l(z fez) = -1 _ l(z

2(z 2(,

+ 1-

Expanding the fractions yields 00

fez)

=-

L n=O

00

z-n-I

+L

2n+ 1z-n- 1

n=O

00

= L(2n+l -

1)z-n-1.

n=O

This is again the unique expansion of f(z) in theregion [z] > 2.

11II

13We could, of course, evaluate the derivatives of all orders of the function at z = 0 anduse Maclaurin's formula. However, thepresent methodgives the sameresultmuchmorequickly.

9.6 TAYLOR ANO LAURENT SERIES

261

9.6.12. Example. Define f (z) as

f

(z) =

(I - cosz)/z2

{ 21

for z '" 0, for z = O.

Wecan show that f(z) is an entire function. Since 1 - cos Z andz2 are entire functions, their ratio is analytic everywhere exceptat thezeros of its denominator. The only suchzerois z = O. Thus,Theorem 9.6.8 implies that f(z) is analyticeverywhere except possibly at z = O. To see the behaviorof f(z) at z = 0, we look atits Maclaurin series: z2n

00

1- cosz = 1- L(-I)n_-, n~O

(2n)!

whichimplies that

!'

The expansion on the RHS shows that the value of the series at z = 0 is which, by definition, is f(O). Thus, the series converges for all z, and Theorem9.6.2 says that f(z) is entire.

III

A Laurent series can give information about the integral of a function around a closed contour in whose interior the function may not be analytic. In fact, the coefficient of the first negative power in a Laurent series is given by a_I =

.2., J f(~) d~. 21f1

rc

(9.17)

Thus, to find the integral of a (nonanalytic) function around a closed contour surrounding zo, we write the Laurent series for the function and read off the coefficient of the I/(z - zo) term. 9.6.13. Example. As an illustration of this idea.Iet us evaluatethe integral I = Pc dz/ [z2(z - 2)], where C is the circle of radius I centeredat the origin.The functionis analytic in the annular region 0 < [z] < 2. We can therefore expand it as a Laurent series about

z = 0 in thatregion:

I I ( I) I 00 z2(z - 2) = - 2z 2 I - z/2 = - 2z 2];

=-~ C2) - ~

(z)n 2:

G) -~ -....

-1,

Thus, L I = and Pc dz/[z2(z - 2)] = 21tia_l = -i1t/2. A directevaluationof the integral is nontrivial. In fact, we will see later that to find certain integrals, it is advantageous to cast them in the form of a contour integral and use either Equation (9.17) or a related equation. II

262

9. COMPLEX CALCULUS

Let f : C -> C be analytic at zoo Then by definition, there exists a neighborhood ofzo in which f is analytic. In particular, we can find a circle Iz-zol = r > a in whose interior f has a Taylor expansion. zero oforder k

9.6.14. Definition. Let f(n) (eo)

00

f(z) = L n=O

I n.

00

(z - zo)" es Lan(Z - zn)". n=O

Then f is said to have a zero oforderk atzo if f(n) (zn) = Oforn = 0, 1, ... , k-l but f(k) (zo) i' o. In that case f(z) = (z - ZO)k. I:~o ak+n(Z- zo)n, where ak We define g(z) as

i' a and [z -zol

< r.

00

g(z) = L

ak+n(z - zo)"

where [z - zol < r

n=O

and note that g(zo) = ak i' O. Convergence of the series on the RHS implies that g(z) is continuous at ZOo Consequently, for each E > 0, there exists 8 such that Ig(z) - akl < E whenever [z - zol < 8.lfwe choose E = lakl/2, then, for some 80> 0, Ig(z) - akl < lakl/2 whenever [z - zo] < 80. Thus, as long as z is inside the circle [z - zn] < 80, g(z) cannot vanish (because if it did the first inequality would imply that lakl < lakl/2). We therefore have the following result. the zeros ofan analytic function are isolated

9.6.15. Theorem. Let f : C -> iC be analytic at zo and f (zo) = O. Then there exists a neighborhood of zo throughout which f has no other zeros unless f is identically zero there. Thus, the zeros ofan analytic function are isolated.

simple zero

When k = 1, we say that zo is a simple zero of f. To find the order of the zero of a function at a point, we differentiate the function, evaluate the derivative at that

point, and continue theprocess untilwe obtain a nonzero valuefor the derivative. 9.6.16. Example. (a) The zeros of cosz, which are z because d coszl = - sin [(2k + dz z~(2k+I)n/2

= (2k + 1),,/2, are all simple,

I)::] '" O. 2

(b) Tofindthe orderof the zeroof f (z) = eZ - 1 - z - z2/2 at z = 0, we differentiate f(z) and evaluate f'(O):

f'(O) = (eZ

-I-z)z=o =

O.

Differentiating again gives f"(O) = (eZ - l)z=O = O. Differentiating once more yields fm (0) = (eZ)z~o = I. Thus, the zero is of order 3. III

9.7 PROBLEMS

263

9.7 Problems 9.1. Show that the function w = l/z maps the straight line y = a in the z-plane onto a circle in the w-plane with radius l/(2Ial) and center (0, l/(2a)). 9.2. (a) Using the chainmle, find af/az* and af/az in terms of partial derivatives with respect to x and y. (b) Evaluate af/az* and af/az assuming that the Cauchy-Riemann conditions hold. 9.3. Show that when z is represented by polar coordinates, the derivative of a function f(z) can be written as

df =e-19 dz

(au +iav), aT aT

where U and V are the real and imaginary parts of f(z) written in polar coordinates. What are the C-R conditions in polar coordinates? Hint: Start with the C-R conditions in Cartesian coordinates and apply the chain IDle to them using x = TCOSO and y = T sinO. 9.4. Show that d/dz(lnz) = liz. Hint: Find u(x, y) and v(x, y) for ln z and differentiate them. 9.5. Show that sin z and cos z have only real roots. 9.6. Show that (a) the sum and the prodnct of two entire functions are entire, and (b) the ratio of two entire functions is analytic everywhere except at the zeros of the denominator. 9,7. Given that u = 2Aln[(XZ + yZ)I/Z], show that v = 2Atan"! (yf x), where u and v are the real and imaginary parts of an analytic function w(z). 9.8. If w(z) is any complex potential, show that its (complex) derivative gives the components of the electric field. 9.9. (a) Show that the flnx through an element of area da of the lateral surface of a cylinder (with arbitrary cross section) is d¢ = dz(IEI ds) where ds is an arc length along the equipotential surface. av/as where v is the imaginary part of the com(b) Prove that lEI Idw/dzl plex potential, and s is the parameter describing the length along the equipotential curves. (c) Combine (a) and (b) to get

=

flux per unit z-length

=

= _¢- = v(Pz) -

V(PI)

zz - Zl

for any two points PI and Pz on the cross-sectional curve of the lateral surface. Conclude that the total flux per unit z-length through a cylinder (with arbitrary

264

9. COMPLEX CALCULUS

cross section) is [v], the total change in v as one goes around the curve. (d) Using Gauss's law, show that the capacitance per unit length for the capacitor consisting of the two conductors with potentials Ut and U2 is c '" charge per unit length _ potential difference -

[v]j4rr

IU2 -

Uti'

9.10. Using Equation (9.7) (a) find the equipotential curves (curves of constant u) and curves of constant v for two line charges of equal magnitude and opposite signs located at y = a and y = -a in the xy-plane. (b) Show that z=a(sin;),. +iSinh;Jj(COSh;),. -cos 2:)

by solving Equation (9.7) for z and simplifying. (c) Show that the equipotential curves are circles in the xy-plane of radii a] sinh(uj2)") with centers at (0, a coth(uj2)")), and that the curves of constant v are circles of radii a] sin(vj2)") with centers at (a cot(vj2)"), 0).

9.11. In this problem, you will find the capacitance per unit length of two cylindtical conductors of radii Rt and R2 the distance between whose centers is D by looking for two line charge densities +),. and -),. such that the two cylinders are two of the equipotential surfaces. From Problem 9.10, we have a

Ri

= sinh(u;j2),.) '

Yi

= a coth(u;j2)"),

i = 1,2,

where Yj and Y2 are the locations of the centers ofthe two conductors on the y-axis (which we assume to connect the two centers). Ut U21 . (a) Show that D = IYt - Y21 = Rt cosh 2)" - R2 cosh 2)" I (b) Square both sides and use cosh(a -b) = cosh a coshb - sinha sinhb and the expressions for the R's and the y's given above to obtain

cosh ( Ut - U2 ) = 2)"

IRr+R~-D21 . 2RtR2

(c) Now find the capacitance per unit length. Consider the special case of two concenttic cylinders. (d) Find the capacitance per unit length of a cylinder and a plane, by letting one of the radii, say Rj, go to infinity while h '" Rj - D remains fixed.

265

9.7 PROBLEMS

9.12. Use Equations (9.4) and (9.5) to establish the following identities. (a) Re(sinz) = sin x coshy,

Im(sin z) = cos x sinh y.

(b) Re(cosz) = cosxcoshy,

Im(cosz) = - sinx sinhy.

(c) Re(sinhz) = sinh x cosy,

Im(sinhz) = cosh x siny.

(d) Re(coshz) = cosh x cosy,

Im(coshz) = sinhxsiny. 2 = cos 2 x + sinh 2 y. 1 coszl

(e) I sind = sin2 x + sinh 2 y, (f) I stnh z]? = sinh 2 x + sin 2 y,

1

coshzl 2 = sinh 2 x + cos 2 y.

9.13. Find all the zeros of sinh z and cosh z. 9.14. Verify the following hyperbolic identities. (a) cosh 2 Z - sinh 2 Z = 1. (b) COSh(ZI + Z2) = COShZl cosh zj + sinh zj sinh zj. (c) sinhtz: + Z2) = sin zj COShZ2 + COShZl sinh z-. (d) cosh2z

= cosh 2 z +

(e) tanhfzj

+ Z2)

sinh 2 z,

sinh2z

= 2 sinh z cosh z,

taoh Zj + tanh Z2

= 1 + tanh zj

taohZ2'

9.15. Show that (a)

taoh(!:) = sinhx+isiny 2

cosh x + cos y

(b) coth

Z) (-2 =

sinhx-isiny . coshx - cos Y

9.16. Find all values of z such that (b) e Z = 1 + i./3.

(a) eZ = -3.

(c)

e

2z- 1

= 1.

9.17. Show that le-zi < 1 if and only if Re(z) > O. 9.18. Show that both the real and the imaginary parts of an aoaiytic function are harmonic. 9.19. Show that each of the following functions-s-call each one u(x, y)-is harmonic, and find the function's harmonic partner, v(x, y), such that u(x, y) + i v (x, y) is analytic. (a)x 3 - 3xi. 2

(d)e- Ycos2x.

(b) eX cosy.

(e) eY

2

2 -X

cos2xy.

(I) eX(x cosy - y siny) + 2 sinh y sin x + x 3

-

3xi + y.

266

9. COMPLEX CALCULUS

9.20. Prove the following identities. (u) cos- I Z = -i In(z ±

(c) tan- I

JZ2 -

I).

I In (i-.-- Z) . z = ---,

,+z

2,

± JZ2 + I).

(e) sinh- I z = In(z

(b) sin- I z = -i In[iz ± ~)].

± ~).

(d) cosh- I

z=

(f) tanh-I

z =! In (I + Z).

In(z

2

I-z

9.21. Find the curve defined by each of the following equations. (u) z = I - it, 0::; t ::; 2. n 3:n: (c) z =u(cost+isint), "2::; t::; 2'

(b) (d)

z = t + it 2 ,

-00

i

z = t + -, t

-00

< t <

00.

< t < O.

9.22. Provide the details of the proof of part (a) of Proposition 9.3.1. Prove part (b) by showing that if f(z)

a2 cf>

a2 cf>

a2 cf> a2 cf>

= z' = x' + iy' is analytic and -ax2 + -ay2 = 0, then

' +--0 axil ay'2 - . 9.23. Let f(t) = u(t) + iv(t) be a (piecewise) continuous complex-valued function of a real variable t defined in the interval a ::; t ::; b. Show that if F(t) = U(t) + i V(t) is a function such that dF[dt = f(t), then

l

b

f(t) dt = F(b) - F(u).

This is the fundamental theorem of calculus for complex variables. 9.24. Find the value of the integral Icr(z + 2)/z]dz, where C is (a) the semicircle z = 2eiO , for 0 ::; it ; (b) the semicircle z = 2e'&, for:n: ::; 2:n:, and (c) thecircle z = 2ei (), for - ] i ~ 8 .:5 Jr.

e ::;

e ::;

9.25. Evaluate the integra1 I y dz/(z-I-i) where y is (a) the line joining zi = 2i and Z2 = 3, and (b) the broken path from zi to the origin and from there to Z2. 9.26. Evaluate the integral Ie zm(z*)ndz, where m and n are integers and C is the circle [z] = I taken counterclockwise. 9.27. Let C be the boundary of the square with vertices at the points

z = I + i, and z = i with counterclockwise direction. Evaluate

fc

(5z +2)dz

and

z = 0, z = I,

9.7 PROBLEMS

267

9.28. Let Cl be a simple closed contour. Deform Cl into anew contour C2 io such a way that C 1 does not encounter aoy singularity of ao aoalytic function f io the process. Show that

J

Tel

J

f(z)dz =

i.

f(z)dz.

That is, the contour cao always be deformed iota simpler shapes (such as a circle) aod the iotegral evaluated.

9.29. Use the result of the previous problem to show that

J

dz.

fCZ- I - 1

= 2,,;

J (z -

aod

fc

I - i)m-1dz

= 0 for m = ±I, ±2, ...

when C is the boundary of a square with vertices at z z = 2;, taken counterclockwise.

= 0, z = 2, z = 2 + 2;, aod

9.30. Use Equation (9.12) aod the bioomial expaosion to show that d

dz (z'") = mzm-l.

9.31. Evaluate '.fc dZ/(Z2 - 1) where C is the circle lzl = 3 iotegrated io the positive sense. Hint: Deform C iota a contour C' that bypasses the singularities of the iotegraod. 9.32. Show that when f is aoalytic witbio aod on a simple closed contour C aod zo is not on C, then

J

fc

f'(z) dz = z - zo

J

f(z) dz .

f c (z -

ZO)2

9.33. Let C be the boundary of a square whose sides lie along the lioes x = ±3 aod y = ±3. For the positive sense of integration, evaluate each of the following integrals. (a)

(d) (g)

(j)

(m)

J

e-

fc z -

i i

z

J

dz:

(b)

i,,/2

sinhz -4-dz.

(e)

c z

cos z

d

c (z - ;,,/2)2

J

Z

2

e

fc z

-5z +4

i

taoz

c

z

(z - a)2

dz

dz:

.

J

eZ

fc Z(Z2 + 10) dz:

(h) (k)

i

coshz

-4-

c z

iJ

c

fc

(e)

dz.

eZ

(z - ;,,)2

dz:

sinhz

~)(Z2 -10)

(f)

-3

(i)

dz:

(I)

(z - ;,,/2)2

for - 3 < a < 3.

(n)

cosz

fc (z _

J

i

i

cos z

c z

d Z.

dz:

cos z

- - . - dz: C Z + in

J

eoshz

fc (z -

dz;

;,,/2)2 Z2

fc (z - 2)(Z2 -

10)

dz:

268

9. COMPLEX CALCULUS

9.34. Let C be the circle lz - i 1 = 3 integrated in the positive sense. Find the valne of each of the following integrals. Z

(a)

(d)

YJ:c Z2 +e rr2 dz • J: dz

(b)

:rc (z2 + 9)2 .

(e)

J: sinhz Yc (Z2 + rr2 j2 dz, J: coshz Yc (Z2 + rr2) 3 dz:

J:

dz

(c)

:rc Z2 + 9'

(f)

J: Z2 Yc z2 -

3z + 4 dz, 4z + 3

9.35. Show that Legendre polynomials (for [x] < I) can be represented as Pn(x) =

(_I)n

i

(1- z2)n

2n(2rri) c (z - x)"

+1 dz;

where C is the unit circle around the origin.

9.36. Let I be analytic within and on the circle Yo given by [z - zol = Yo and integrated in the positive sense. Show that Cauehy's inequality holds: nlM

1I(n)

(zo)1 :::: ~, Yo

where M is the maximum value of I/(z)1 on Yo.

9.37. Expand sinhz in a Taylor series about the point z = in . 9.38. What is the largest circle within which the Maclaurin series for tanh Z converges to tanh z? 9.39. Find the (unique) Laurent expansion of each of the following functions about the origin for its entire region of analyticity. (a)

I (z - 2)(z - 3)

I

.

I

I (f) ---Z--I'

(e) (I _ z)3' (i)

(d) sinhz-z

(e) --'z2'--(-'-I---z-'--).

z -

Z4

z2 -4 (g) ---Z--9'

(h) (Z2

z -

~ 1)2'

_z_. z-I

9.40. Show that the following functions are entire. (a) I(z) =

{

e2Z _ 1 2 -----;:- -

for Z f= O,

2

for z = O.

z

cos z (e) I(z) =

Z2 - rr

{ -I/rr

2

/4

for z

f= ±rr/2,

for z =

±rr/2.

(b) I(z) =

{

~

I z

for z f= 0, for z = O.

9.7 PROBLEMS

9.41. Let I be analytic at zo and I (zo) = I' (zo) that the following function is analytic at zo:

I(z)

269

= ... = I(k) (eo) = O. Show

for z i' zo.

(z - ZO)k+l g(z) =

1+

flk+ll(zo) (k I)!

for z = zoo

9.42. Obtain the first few nonzero terms of the Laurent series expansion of each of the following functions about the origin. Also find the integral of the function along a small simple closed contour encircling the origin.

I sm z

(b)

(a) -.-. Z4

(e)

6z+z

I 1- cosz

3

-6sinhz

.

(f)

I -2-'-' Z sm z

(c)

z

I-coshz I (g) e Z - I'

(d) - - , - -

z - sin z

Additional Reading I. Churchill, R. and Verhey, R. Complex Variables and Applications, 3rd ed., McGraw-Hill, 1974. An introductory text on complex variables with many examples and exercises. 2. Lang, S. Complex Analysis, 2nd ed., Springer-Verlag, 1985. An extremely well-written book by a master expositor. Although the book has a formal tone, the clarity of exposition and the use of many examples make this book very readable.

10

_

Calculus of Residues

One of the most powerful tools made available by complex analysis is the theory of residues, which makes possible the routine evaluation of certain definite integrals that are impossible to calculate otherwise. The derivation, application, and analysis of this tool constitute the main focus of this chapter. In the preceding chapter we saw examples in which integrals were related to expansion coefficients of Laurent series. Here we will develop a systematic way of evaluating both real and complex integrals.

10.1 Residues

isolated singuiarity

Recall that a singular point zo of f : iC --* iC is a point at which f fails to be analytic. If in addition, there is some neighborhood of zo in which f is analytic at every point (except of course at zo itself), then zo is called an isolated singuIarity of f. Almost all the singularities we have encountered so far have been isolated singularities. However, we will see later-when discussing multivalued functions-that singularities that are not isolated do exist. Let zo be an isolated singularity of f. Then there exists an r > 0 such that within the "annular" region 0 < [z - zo I < r, the function f has the Laurent

expansion f( Z) =

~

( n ~ n L- anz-zo) =L-an(Z-ZO)

n=-oo

n=O

where

1 an = 2".;

J

f(~)d~

fc (~ - zo)n+l

and

bl b2 +--+ )2+'" Z - zo (z - zo

10.1 RESIDUES

271

In particular,

1 hi = 2 . 7T:Z

residue defined

1. f(~) d~, Yc

(10.1)

where C is any simple closed contour around zo. traversed in the positive sense, on and interior to which f is analytic except at the point zo itself. The complex number hj, which is essentially the integral of f(z) along the contour, is called the residue of f at the isolated singular point zo. It is important to note that the residue is independent of the contour C as long as zo is the only isolated singular point within C.

Pierre Alphonse Laurent (1813-1854) gradoated from the EcolePo1ytechniqoe near the top of his classand became a secondlieutenant in theengineering corps. Onhisreturn from the war in Algeria, he took part in the effort to improve the portat Le Havre, spending six years there directing various parts of theproject Laurent's superior officers admired the breadth of his practical experience andthe good judgment it afforded the young engineer. During this period he wrote his first scientific paper, on the calculus of variations. and submitted it to the French Academy of Sciences for the grand prix in mathematics. Unfortunately the competition hadalready closed(allbough thejudgeshadnot yetdeclared a winner), andLaurent's submission was not successful. However, thepaper so impressed Cauchy that he recommended its publication, also without success. Thepaper for which Laurentis mostwellknown suffered a similar fate. In ithedescribed amore general form of atheorem earlier provenby Cauchy for the power series expansion of afunction. Laurentrealizedthat onecouldgeneralize thisresult to holdinany annular region between two singular or discontinuous points by usingbothpositive andnegative powers in the series, thus allowing treatment of regions beyond thefirst singular or discontinuous point. Again, Cauchy argued for thepaper's publication without success. Thepassage of time provided a more justreward, however, andtheuse of Laurent series became a fundamental toolin complex analysis. Laurent later worked in thetheory of lightwaves and contended withCauchy overthe interpretation of thedifferential equations thelatter hadfonnulated to explain thebehavior of light. Little came of his work in this area. however, andLaurent died attheageof fortytwo,_a captain serving on the committee on fortifications in Paris. His widow pressed to havetwo more of hispapers read to theAcademy, only one of whichwaspublished.

We use the notation Res[f(zo)] to denote the residue of f at the isolated singular point zoo Equation (10.1) can then be written as

t

f(z) dz = 2ni Res[f(zo)]·

What if there are several isolated singular points within the simple closed contour C? The following theorem provides the answer. residue theorem

10.1.1. Theorem. (the residue theorem) Let C be a positively oriented simple

272

10. CALCULUS OF RESIDUES

Figure 10.1

Singularitiesare avoidedby going around them.

closed contour within and on which a function f is analytic except at a finite number of isolated singular points Zl , Z2, ... , Zm interior to C. Then

1.

fc

f(z)dz=2,..itReS[f(Zk)].

(10.2)

k=1

Proof Let Ck be the positively traversed circle around Zk. Then Figure 10.1 and the Cauchy-Goursat theorem yield

f

0= !cJ(Z)dZ = -

f(z)dz

f

+

circles

f(z)dz +!c f(z)dz,

parallel lines

where C' is the union of all the contours, and the minus sign on the first integral is due to the interiors of all circles lie to our right as we traverse their boundaries. The contributious of the parallel lines cancel out, and we obtain

i

c

f(z) dz

=t k=l

i

f(z) dz

=t

Ck

2,..i Res[f(Zk)],

k=l

where in the last step the definition of residue at Zk has been used.

o

10.1.2. Example. Let us evaluate the integral :fc(2z - 3) dz/[z(z - 1)] where C is the circle Izi = 2. There are two isolated singularities in C, Zl = 0 and Z2 = I. To find Res[f(ZI)l. we expand around the origin: 2z-3 3 I 3 --- = - - -- = z(z - I) z z- I z

I

3

+- = -z + 1 + z +... 1- z

for [z] < I.

10.2 CLASSIFICATION OF ISOLATED SINGULARITIES

This gives Res[!(z,)]

273

= 3. Similarly, expandingaroundz = I gives

2z-3 = 3 _ _I_= _ _I_+ 3I:(_I)n(Z_I)n, z(z - I) z- I+ I z- I z- I k=O

whichyields Res[!(z2)] = -I. Thus,

1. 2z Yc z(z -

3 dz = 2"iIRes[!(z,)] + Res[f(Z2)]} = 2"i(3 - I) = 4"i.

III

I)

10.2 Classification of Isolated Singularities Let [ : iC -> iC have an isolated singularity at zoo Then there exist a real number r > 0 and an annular region 0 < [z - zol < r such that T can be represented by

the Laurent series (10.3)

principal partofa function

removable singular point

The second sum in Equation (10.3), involving negative powers of (z - zo), is called the principal part of [ at zoo We can use the ptincipal part to distinguish three types of isolated singularities. The behavior of the function near the isolated singularity is fundamentally different in each case.

I. If bn = 0 for all n 2: I, zo is called a removable singular point of f . In this case, the Laurent series contains only nonnegative powers of (z - zo), and setting [(zo) = ao makes the function analytic at zoo For example, the function fez) = (e' - 1 - Z)/z2, which is indeterminate at z = 0, becomes entire if we set [(0)

= 1,because its Lanrentseries fez) = ~+~+:~ +...

has no negative power. poles defined

2. If bn = 0 for all n > m and bm # 0, zn is called a pole of order m. In this case, the expansion takes the form

~ n [(z) = L.., an (z - zn) n~O

simple pole essential singularity

for 0 <

lz -

»: + -bi- + ... + -:---=-:-:::Z -

zo

(z - zo)"

zn] < r. In particular, if m = 1, ZO is called a simple pole.

3. If the principal part of [ at zo has an infinite number of nonzero terms, the point zo is called an essential singularity. A prototype of functions that have essential singularities is

274

10. CALCULUSOFRESIOUES

which has an essential singularity at

z=

a and a residue of I there. To see

how strange suchfunctions are, we let a be any realnumber, and consider z = I/Ona + 2n,,0 for n = 0, ±1, ±2, .... For such a z we have e 1/ z = elna+2mri = ae 2nni = a. Inparticular, as n ---+ 00, Z gets arbitrarily close to the origin. Thus, in an arbitrarily small neighborhood of the origin, there are infinitely many points at which the function exp(l/z) takes on an arbitrary value a. In other words, as z ---> 0, the function gets arbitrarily close to any real number! This result holds for all functions with essential singularities.

10.2.1. Example. ORDER OF POLES (a) The function (zz - 3z +5)/(z - I) hasa Laurent series around z = I containingonly 1hreetenns: (zZ -3z+5)/(z -I) = -I + (z-I)+3/(z -I). Thus, it has a simplepole at z = 1, with a residue of 3. (b) The function sinz/z6 has a Laurent series sinz

7

1 00 z2n+l 1 lIz = z6 ~(_I)n (2n + 1)1 = z5 - 6z3 + (5!)z -

7! + ...

about z = O. Theprincipal part has three terms. The pole, at z = 0, is of order 5, and the function has a residue of 1/120 at z = O. (c) The function (zZ - 5z + 6)/(z - 2) has aremovahle singularity at z = 2, because

zZ - 5z +6 = (z - 2)(z - 3) = z _ 3 = -I + (z -2) z-2

andbn =

z-2

a for all n.

III

10.2.2. Example. SINGULARITIES OF A RATIONAL rational function

FUNCTION

In this example we show that a function whose only singularities in the entire complex

plane arepoles musthe a rational function, 1 i.e., the ratio of two polynomials. Let f be suchafunction and let {z j lJ=l be its polessuchthat Z j is of order m l : Expand thefunction about zt in a Laurent series bj fez) = - Z -

Afunction whose only singularities in the entire complex plane are poles is a rational function.

Zl

+ ... + (Z -

bmj Zl

)mj

~

+ L" ak(Z k=O

Zt)

k

== (

Pt(z) )mj

Z - Zl

+ 81(Z),

where Pj (z) is a polynomialof degree mj - I in z and 81 is analytic at Zj. It should be clearthat the remaining poles of f are in gl. So, expand gl about Z2 in a Laurent series. A similar argumentas above yields 81 (z) = Pz(z)/(z - Zz)m2 + 8Z(Z) where P2(Z) is a polynomial of degree m2 - I in z andg2 is analytic at zt and Z2. Continuing in thismanner, we get

where g hasno poles. Sinceallpoles of f havebeenisolatedin the sum, g mustbe analytic everywhere in C, i.e., an entire function. Now substitute 1/ t for z, take the limit t --+ 0, lWe assume that thepointatinfinity is not a pole of the function, i.e., that f(l/z) does not havea pole at theorigin.

10.3 EVALUATION OF DEFINITE INTEGRALS

andnote that, since thedegreeof Pi is mi to zeroexceptpossibly gO/t). Moreover, lim g(l/t) HO

-

275

1, all the terms in thepreceding equation go

i' 00,

because, by assumption, thepointat infinity is not a pole of f. Thus, g is a bounded entire function. By Proposition 9.5.5, g mustbe a constant. Taking a common denominator forall the terms yields a ratioof two polynomials. II

The type of isolated singularity that is most important in applications is of the second type-poles. For a function that has a pole of order m at zo, the calculation of residaes is routine. Such a calculation, in turn, enables us to evaluate many integrals effortlessly. How do we calculate the residue of a function f having a pole of order m at zo? It is clear that if f has a pole of order m, then g : iC --> iC defined by g(z) es (z - zo)m f(z) is analytic at zoo Thus, for any simple closed contour C that contains zo but no other singular point of f, we have Res[f(zo)]

I i =i i

= -2' 1l'1

C

-2' 7n

f(z)dz

C

g(z)dz'

(z - zo)m

g(m-l)(zo) 1)1' .

= (m -

In terms of f this yields 2 I Res[f(zo)] = (

m-

d m- I 1)1 lim d .. -I [(z - zo)" f(z)].

. z-e-zn

Z

(lOA)

For the special, but important, case of a simple pole, we obtain

Res[f(zo)] = lim [(z - zo)f(z)].

(10.5)

Z--+ZQ

10.3 Evaluation of Definite Integrals The most widespreadapplication ofresidues occurs in the evaluation ofreal definite integrals. It is possible to "complexify" certain real definite integrals and relate them to contour integrations in the complex plane. We will discuss this method shortly; however, we first need a lemma.

10.3.1. Lemma. (Jordan's lemma) Let CR be a semicircle ofradius R in the upper half of the complex plane (UHP) and centered at the origin. Let f be a function that tends uniformly to zero faster than l!I'zl for arg(z) E [0, n] as [z] --> 00. Let 0: be a nonnegative real number. Then

2Thelimitis taken becausein manycases themeresubstitution of zo mayresultin anindeterminate form.

276

10. CALCULUS DF RESIDUES

Proof. For Z E C R we write z = Rei9, dz = I Re i9de, and IOIZ = IOI(R cos e

+ IR sin e) =

IOIR cos e - OIR sine

and substitute in the absolute value of the integral to show that IIRI

:s1o" e-·Rsin9 Rlf(Re i9)lde.

By assumption, Rlf(Re i9)1< feR) independent of e, where feR) is an arbitrary positive number that tends to zero as R ---> 00. By breaking up the interval of integration into two equal pieces andchanging e to n - e in the second integral, one can show that IIRI < 2f(R) 10,,/2 e-· Rsin9de.

"=

Furthermore, sin e Thus,

Ihl

< 2f(R)

2e In for 0

" f2

foo

:s e :S n 12 (see Figure 10.2 for a "proof").

. nf(R) e-(2.R/,,)9de = - - ( I - e-· R), OIR

which goes to zero as R gets larger and larger.

D

Note that Jordan's lemma applies fora = 0 as well, because (I-e-· R) ---> OIR as a ---> O. If a < 0, the lemma is still valid if the semicircle CR is taken in the lower half of the complex plane (LHP) and fez) goes to zero mtiformly for n :S arg(z) :S 2n. We are now in a position to apply the residue theorem to the evaluation of definite integrals. The three types of integrals most commonly encountered are discussed separately below. In all cases we assume that Jordan's lemma holds.

10.3.1 Integrals of Rational Functions The first type of integral we can evaluate using the residue theorem is of the form

-100

I 1-

-00

p(x) - dx q(x)

,

where p(x) and q(x) are real polynomials, and q(x) then write

t, = lim

1

R

p(x) dx R....oo -R q(x)

=

f=

0 for any real x. We can

r

lim p(z) dz R....oo lex q(z) ,

where Cx is the (open) contour lying on the real axis from - R to +R. Assuming that Jordan's lemma holds, we can close that coutour by adding to it the semicircle

10.3 EVALUATION DF DEFINITE INTEGRALS

277

....__...i···········-r-·········-t·__ ······_+·····--T .........i···········_j_···········f··

I

_·_·-t-·_---------:--·_···_----+----

·····1

----------r-------A ---+-------!----------'\.

----r-··.--

-!

1/,,-

.i.,

-+.--------.

I

1.5

!--------~---'\i,--------

o -0.5

0

0.5

2

2.5

3

3.5

Figure 10.2 The "proof" of sin 6 ::: 26/" for 0", 6 '" "/2. The line is the graph of y = 26/,,; the curve is that of y = sin 6. of radius R [see Figure lO.3(a)]. This will not affect the value of the integral, because in the limit R --+ 00, the contribution of the integral of the semicircle tends to zero. We close the contour in the UHP if q(z) has at least one zero there. We then get

!J = lim

J:

~Res [P(Zj)]

p(z) dz = 2"j R-+C.,rCq(z) ~

q(Zj)

,

where C is the closed contour composed ofthe interval (- R, R) and the semicircle C R, and (Zj 11~t are the zeros of q(z) in the UHP. We may instead close the contour in the LHP,3 in which case

where (Zj I1=1are the zeros of q(z) in the LHP. The minus sign indicates that in the LHP we (are forced to) integrate clockwise.

10.3.2. Example. Let us evaluate the integrall =

1000 x 2 dx/[(x 2 + 1)(x2 + 9)]. Since

the integrand is even, we can extend the interval of integration to all real numbers (and dividethe result by 2). It is shownbelowthat Jordan's lemma holds. Therefore, we write the contour integral corresponding to I:

3Provided thatJordan's lemmaholdsthere.

278

10. CALCULUS OF RESIDUES

where C is as shown in Figure 1O.3(a). Note that the contour is traversed in the positive sense.This is alwaystrue fortheUHP. Thesingularities of thefunction in theUHPare the simple poles i and 3i corresponding to the simple zeros of the denominator. The residues atthesepoles are 2 Res[f(i)] = lim (z _ i) z z....t (z - i)(z + i)(z2 Res[f(3i)]

=

lim (z _ 3i) z....3'

1 --

+ 9)

z2 (z2 + l)(z - 3i)(z

16i '

+ 3i)

=

3 16i'

Thus,we obtain

It is instructive to obtain the sameresults using the LHP. In thiscase, the contour is as shownin Figure 10.3 (b) andis taken clockwise, so we haveto introduce a minussign.The singular points areat z = -i and z = - 3i. Thesearesimplepoles at whichtheresidues of thefunction are 2

Res[f(-O] = lim (z + O-:----::--,----,-z--,---,,---:::z....-,

Res[f(-3i)] =

(z - i)(z

lim (z + 3i)

z....-3'

+ i)(z2 + 9)

16i'

2

(z2

z

+ 1)(z -

3i)(z

+ 3i)

=

3 16i

Therefore,

To show thatJordan's lemma applies to this integral, we have only to establish that limR-4oo Rlf(Rei8)1 = O. In the case at hand, a = 0 because there is no exponential function in the integrand. Thus,

whichclearlygoes to zeroas R

--+ 00.

11II

10.3.3. Example. Let us uow consider a slightly more complicated integral:

1

00

-00

x2dx (x 2 + 1)(x 2 + 4)2'

Pc

which in the UHP turns into z2 dz/[(z2 + l)(z2 + 4)2]. The poles in the UHP are at z = i and z = 2i. The former is a simple pole, andthe latter is a pole of order 2. Thus,

10.3 EVALUATION OF DEFINITE INTEGRALS

279

3i

R

-R

-i -3i

(a)

(b)

Figure 10.3 (a) The large semicircleis cbosen in the UHP. (b) Note bowthe directionof contour integration is forced to be clockwisewhenthe semicircle is chosen in theLHP.

using Equations (l0.5) and (10.4),we obtain Res[f(i)l= lim(z-i)

z....i

z2

(z - i)(z

+ i)(z2 + 4)2

.'!- [(Z _ 2i)2

Res[f(2i)] = _ 1 _ lim (2 - 1)1 z....2' dz d [

=

,~, dz

= __1_, 18i

z2 ] (z2 + I)(z + 2i)2(z - 2i)2

z2 ] 5 (z2 + I)(z + 2i)2 = 72i'

and

1_ + _5_) _ z.

oo ~~.::x_2=dx:;.-_~ _ 2"i ( __ -00 (x 2+1)(x2+4)2 18i

f

72i

- 36'

Closingthecontour in theLHP wouldyield thesameresult.

III

10.3.2 Products of Rational and Trigonometric Functions The second type of integral we can evaluate using the residue theorem is of the . form 00 pIx)

1

--cosaxdx

-00 q(x)

or

1

00

pIx) .

--smaxdx,

-00 q(x)

where a is a real number, pIx) and q(x) are real polynomials in x, and q(x) has no real zeros. These integrals are the real and imaginary parts of I 2=

1

pIx) 'ax d x. -00 q(X) 00

--e

280

10. CALCULUS DF RESIDUES

The presence of e iax dictates the choice of the half-plane: If a "=: 0, we choose the UHP; otherwise, we choose the LHP. We must, of course, have enough powers of x in the denominator to render Rjp(Reie)/q(ReiB)1 uniformly convergent to zero. 10.3.4. Example. Let us evaluate!-""oo[cos ax/(x 2 + IP] dx where a 'i O. This integral is the real part of the integral h = J~oo e iax dx / (x 2 + 1)2. When a > 0, we close in the UHP as advised by Jordan's lemma. Then we proceed as for integrals of rational functions. Thus, we have

h =

i

eiaz

C (z

2

+ I)

2 dz

= 2rri Res[f(i)]

fora> 0

because there is only one pole (of order 2) in the UHP at z = i, We next calculate the residue:

Res[f(i)] = lim .'!..z-->i dz

. = Inn

-d z-+i dz

[(z - i)2 (z - i)e~az(z + i) 2J iaz

[ -e- (z + i)2

J= Inn. [CZ + i)iaeiaz - 2eiaZJ = -e-(a I +a). z-vt

(z

Substituting this in the expression for 12. we obtain

h

+ i)3

= Ie-a(l

4i

+ a) for a >

O.

When a < 0, we have to close the contour in the LHP, where the pole of order 2 is at z = -i and the contour is taken clockwise. Thus, we get

h=

eiaz 2 2 dz C (z + I)

i

= -2rri Res[t(-i)]

for a < O.

For the residue we obtain

iaz

e2 Res[f(-i)] = lim -d [ (z + i)2 z-->-i dz (z - i) (z and the expression for

h

becomes

h

=

~ea(l

+ i)

2

a J= --(1e a), 4i

- a) for a < O. We can combine the two

results and write

1

00

-00

cosax 2 2 dx

(x

+ I)

7f

I I

= Re(l2) = lz = -(I + lal)e- a . 2

III

10.3.5. Example. As another example, let us evaluate

1

00

-00

x sin ax ---dx x4 + 4

where a 'i O.

This is the imaginary part of the integral h = J~oo xeiax dx/(x 4 +4), which, in terms of z and for the closed contour in the UHP (when a > 0), becomes

tx>

i

zeiaz m -4--dz=2rriLRes[f(Zj)]

C z +4

j~l

for a> O.

(10.6)

10.3 EVALUATION OF DEFINITE INTEGRALS

281

The singularities are determined by the zeros of the denominator: z4 + 4 = 0, or Z = 1 ± i, -1 ± i, Of these four simplepoles only two, 1 + i and-1 + i, are in the UHP. We now calculate theresidues: zei az Res[f(1 + i)] = lim (z - I - i) . . . . z-->1+; (z - I -,)(z - I + ,)(z + I-,)(z + 1+,) (1 + i)e ia(l + i) eiae- a

(2i)(2)(2 Res[f(-I + ill =

lim

z-->-1+;

(-1

+ 2i)

(z + I - i)

+Oia (-

8i ze iaz . . . . (z + I -,)(z + I + ,)(z - I-,)(z - 1+,)

l + i)

(2i)(-2)(-2 + 2i)

e-iae- a

8i

Substituting in Equatiou(10.6),we obtain e- a . . x 12 = 2rci""""8i(ela _e-W ) = ize-asina.

Thus,

1

00

-00

x sin ax 1r -4-- dx = lm(h) = _e- a sin a x +4 2

for a> O.

(10.7)

For a < 0, we could close the contour in the LHP. But there is an easierway of getting to the auswer. We note that -a > 0, aud Equation (10.7)yields

1

00

-00

1

00

xsinax x4 + 4

--dx=-

-00

xsin[(-a)x] 1r (a). n dx=--e-- sm(-a)=-easina. x4 + 4 2 2

Wecancollect thetwo cases in

1

00

-00

x sin-ax ---dx x4 + 4

= 1C_e- Ia I sina. 2

10.3.3 Functions of Trigonometric Functions The third type of integral we can evaluate usingthe residue theorem involves only trigonometric functions and is of the form

fo2n F(sinO, cos 0) ao. where F is some (typically rational) function of its arguments. Since 0 varies from

oto 217: , we can consider it an argument of a point z on the unit circle centered at the origin. Then z = eW and e-w = I/z, and we can substitute cos 0 = (z + l/z)/2, sinO

J

= (z -

Jc

F

l/z)/(2i), and dll

= dz/(iz) in the original integral, to obtain

(Z - I/z, z + I/Z) ~z. 2,

2

IZ

This integral can often be evaluated using the method of residues.

282

10. CALCULUS OF RESIDUES

10.3.6. Example. Let us evaluate the integral Substituting for cos () and

fr

dO/(1 + acosO) where faJ < 1.

de in terms of z, we obtain 21. dz

1. dz/iz f c 1+ a[(z2 + 1)/(2z)] =

i

fc 2z + az 2 + a'

where C is the unit circle centered at the origin. The singularities of the integrand are the

zeros of its denominator: ai =

-l+~

and

a

-I-~

Z2 =

a

For lal < 1 it is clear that Z2 will lie outside the unit circle C; therefore, it does not contribute to the integral. But zt lies inside, and we obtain

fc

2z + ::2 + a = 2:n:i Res[f(Zt)]·

The residue of the simple pole at zt can be calculated:

Res[!(ZI)]

= z-e-z lim (z t

Z2

)

a) = 2v'I-a2'

I (

=;;

Zl) ( ~( a Z - zr Z -

~ (_1_) a zt - Z2

I

2v'I-a 2

It follows that

[2"

10

2 1.

dO

2

dz

(

I

)

1+{lcose=irc2z+az2+a=i'bri 2-!1-a2

2:n: =

Jl-a Z'

II

10.3.7. Example. As anotherexample, let us consider the integral ] _ {"

dO

- 10 (a + cos 0)2 where a> 1. Since cos e is an even function of e. we may write ]-

11"

-2:

dO _,,(a+cosO)2

where a > 1.

This integration is over a complete cycle around the origin, and we can make the usual substitution:

]=~1. dz/iz =~1. zdz . i fc (z2 +2az + 1)2 2 f c [a + (z2 + 1)/2z]2

J

vi

The denominator has the"roots zt = -a + a Z - 1 and Z2 = -a a 2 - 1, which are both of order 2. The second root is outside the unit circle because a > 1. Also, it is easily verified that for all a > 1, zr is inside the unit circle. Since z t is a pole of order 2, we have

Res[f(ZI)] = lim

'!. .- [(z _ ZI)2 (z -

z-->z[ dz

-lim'!...-[ - z-e-zr dz

We thos obtain]

z

]_

z

ZI)2(z - Z2)2

I

(z - Z2)2 - (Zl - Z2)2

= ~2:n:i Res[!(ZI)] = 2:n:a 3/2' , (a - I)

]

10.3 EVALUATION OF DEFINITE INTEGRALS

283

10.3.4 Some Other Integrals The three types of definite integrals discussed above do not exhaust all possible applications of the residue theorem. There are other integrals that do not fit into any of the foregoing three categories but are still manageable. As the next two

examples demonstrate, aningenious choice of contours allowsevaluation of other types of integrals.

10.3.8. Example. Let us evaluate the Gaussianintegral I =

1

00

.

2

e lax-bx dx

where a, b

E

lll., b > O.

-00

Completing squares in the exponent, we have 1=

1

00

e- b[x-ia/(2b)]2_a2j4b d x = e- a2/ 4b

-00

lim

l

R

e- b[x-ia/(2b)]2 dx.

R-?oo -R

If we change the variable of integration to z = x - iaf(2b), we obtain 1= e -2/(4b) alim

l

R - ia / (2b)

e?bZ2 dz.

R-+oo -R-iaj(2b)

Let us now define IR: [R es

l

R - ia / (2b)

2

e- bz dz.

-R-ia/(2b)

This is an integralalong a straightline Ci that is parallel to the x-axis (see Figure 10.4). We close the contour as shown and note that e- hzZ is analyticthroughout the interior of the closed contour (it is an entirefunction!). Thus, the contour integral must vanishby the Cauchy-Gaursat theorem. So we obtain IR + [

e- bz2 dz

+

i:

e- bx2 dx

R

C3

Along C3, z = R

+

1

e- bz2 dz

=

O.

C4

+ iy and

r e-bz2dz=1°

lc,

e-b(R+iy)2idY=ie-bR21°

-ia/(2b)

eby2-2ibRYdy

-ia/(2b)

which clearly tends to zero as R --+ 00. We get a similarresultfor the integral along C4. Therefore, we have IR =

l

R

2

e- bx dx

-R

=}

.

11m IR =

R~oo

1

00

-00

2

e _bx dx =

~ -b'

Finally, we get III

284

10, CALCULUS OF RESIDUES

-R

R

-ia 1(2b) Figure 10.4

CI

The contour for the evaluation of the Gaussian integral.

10.3.9. Example. Let us evaluate [ fooo dxl(x 3 + I). If the integraud were even, we could extend the lower limit of integration to -00 and close the contour in the UHP. Since this is not the case, we need to use a different trick. To get a hint as to how to close the contour, we study the singularities of the integrand. These are simply the roots of the denominator: z3 = -1 or Z,~ = e i (2n + l )1l' j 3 with n 0, 1,2. These, as well as a contour that has only zo as an interior point, are shown in Figure 10.5. We thus have

=

=

[+

1

dz -3-+I

+

CR Z

1c,

dz -3-= 2"i Res[f(zo)]· z +I

(10.8)

The C R integral vanishes, as usual. Along Cz. z = ria, with constantc, so that dz aud

1

dz C2 z3 + 1 =

1 0

00

ei'dr ,{OO dr UX (re ia)3 + 1 = _e r 3e 3ia

io

+ 1.

In particular, if we choose 3a = 2rr. we obtain

f ~=

1c, z3 + I

1 ~ = _i 27t/ 3 I. 3 00

_i21r/3

10 r + I

Substituting this in Equation (10.8) gives

(I - e i2rr/3)[

=

2,,; Res[f(zO)]

=}

[=

21Ci

'2 /3 Res[f(zo)],

1- e'

71:

On the other baud, Res[f(zO)] =

lim (z - zo) 2-->20

I

I (z - zo)(z - ZI)(Z - Z2) I

(zo - ZI)(ZO - Z2)

(e irr/3 - eirr)(eirr/3 _ e i5rr/3) '

These last two equations yield

2"i

I

[ = 1- e i2rr/3 (e irr/3 _ eirr)(eirr/3 - ei5rr/3) =

27r 3,,13'

= ei a dr

285

10.3 EVALUATION OF DEFINITE INTEGRALS

Cz

• Zz Figure 10.5

The contour is chosenso that only one of the poles lies inside.

10.3.5 Principal Value of an Integral So far we have discussed only integrals of functions that have no singularities on the contour. Let ns now investigate the conseqnences of the presence of singular points on the contour. Consider the integral

1'"

f(x) dx,

(10.9)

_ooX-XQ

where xo is a real number and f is analytic at xo. To avoid xo-which causes the integrand to diverge-we bypass it by indenting the contour as shown in Figure 10.6 and denoting the new contour by Cu' The contour Co is simply a semicircle of radius E. For the contour Cu , we have

(

f(z) dz =

Jeu z principal value of an integral

XQ

1

xo

-<

-00

f(x) dx x - XQ

+

1'"

f(x) dx

xo+€ X -

+(

f(z) dz,

leo z -

XQ

XQ

In the limit E ~ 0, the sum of the first two terms on the RHS-when it existsdefines the principal value of the integral in Equation (10.9):

P

1'"

-00

f(x) . --dx=lim x - XQ E~O

[1

xo

-<

-00

-f(x) -dx+ X -

XQ

1'"

-f(x) - d x].

xo+€ X

-

XQ

The integral over the semicircle is calcnlated by noting that z - Xo = .e;e and dz = iee"dO: JeD f(z) dz/(z - xo) = -i:n:f(xo). Therefore,

1

f(z) - dz = P

CuZ-XQ

1'"

f(x) . - dx - mf(xo)·

_ooX-XQ

(10.10)

286

10. CALCULUS OF RESIDUES

Cu Xo-E

XO+E

The contour Cu avoids xo.

. Figure 10.6

On the other hand, if Co is taken below the siogularity on a contour Cd, say, we obtaio

1

00

f(z) dz = P

(

Jed Z - XQ

-00

f(x) dx

+ irrf(xo).

x - XQ

We see that the contour integral depends on how the siogular poiot Xo is avoided. However, the priocipal value, if it exists, is unique. To calculate this priocipal value we close the contour by addiog a large semicircle to it as before, assuming that the contribution from this semicircle goes to zero by Jordan's lemma. The contours C« and Cd are replaced by a closed contour, and the value of the iotegral will be given by the residue theorem. We therefore have

1

00

P

-00

f(x) dx = ±irrf(xo) + 2rri fRes [ f(Zj) ],

x-

XQ

j=l

(10.11)

Zj - XQ

where the plus sign corresponds to placing the iofioitesimal semicircle io the UHP, as shown io Figure 10.6, and the mious sign corresponds to the other choice. 10.3.10. Example. Let us use the principal-value method to evaluate the integral 1=

11

00 sinx sinx --dx = --dx. ox 2_ 00 x 00

10

It appears that x = 0 is a singular point of the integrand; in reality, however, it is only a removable singularity, as can be verifiedby the Taylor expansionof sinxJx. To makeuse of theprincipal-value method, we write

I = ~2 1m (1

ix

00

e

-00

x

dX) = ~2 1m (p 1

ix

00

-00

e

x

dX).

We now use Equation (10.11) with the small circle in the UHP, noting that there are no singularities for e ix [x there. This yields 00

P

1

-00

ix

-

x

dx

= breeD)

= in.

Therefore,

roo sinx

10 7

1

dx =

1r

2: Imuz) = 2'

III

287

10.3 EVALUATION OF DEFINITE INTEGRALS

I

e _ _ _ _1

- - - - -



x

Xo

Figure 10.7 10.6.

The equivalent contour obtained by "stretching" Cu , the contour of Figure

The principal value of an integral can be written more compactly if we deform the contour Cu by stretching it into that shown in Figure 10.7. For small enough <, such a deformation will not change the number of singularities within the infinite closed contour. Thus, the LHS of Equation (10.10) will have limits of integration -00 + i< and +00 + i<.1f we change the variable of integration to ~ = z - ie, this integral becomes

1

00

-00

f(~ + i<) d~ ~ + ie - XQ

=

1

00

-00

f(~) d~ ~ - XQ + ie

=

1

00

-00

f(z) dz Z - XQ +

ie'

(10.12)

where in the last step we changed the dummy integration variable back to z. Note that since f is assumed to be continuous at all points on the contour, f (~ + i s) ---> f(~) for small <. The last integral of Equation (10.12) shows that there is no singularity on the new x-axis; we have pushed the singnlarity down to XQ - i <. In other words, we have given the singnlarity Onthe x-axis a small negative imaginary part. We can thus rewrite Equation (10.10) as

1

00

p

-00

f(x) dx = irrf(xQ)

x-

+

XQ

1

00

-00

f(x) dx. , x - XQ + lE

where x is used instead of z in the last integral because we are indeed integrating along the new x-axis-s-assumlng that no other singnlarities are present in the UHP. A similar argument, this time for the LHP, introduces a minus sign for the first term on the RHS and for the e term in the denominator. Therefore,

1

00

p

-00

f(x) dx = ±irrf(xQ) +

x - xo

1

00

-00

f(x)dx., x - XQ ± lE

(10.13)

where the plus (minus) sign refers to the UHP (LHP). This result is sometimes abbreviated as

1

I

-----,--.- = P - - dx Of x -XO±ZE x -XD

irr~(x

-

XQ).

(10.14)

288

10. CALCULUS DF RESIDUES

10.3.11. Example. Let us use residues to evaluatethe fuuction

1

The integral representation 01 the e (step) lunction

ikx

00

I f(k)=-. 21n

e - dx .,

E

>

-00 X - l f

o.

Wehaveto close thecontour by adding a large semicircle. Whether we do this in theUHP or the LHP is dictatedby the sign of k: If k > 0, we close in the UHP. Thus, 1 = -.

f(k)

21l'l

1- - . = eikZdz

[ e

C Z - IE

[(Z -

= lim

it:)

Z-+l€

Z -

eik~

Z -

i kz

Res - - . ]

IE

]

z-vie

= e-kE-----+ 1.

IE

10-+0

On the other hand, if k -c 0, we mustclose in the LHP, in whichthe integrand is analytic. Thus, by theCauchy-Goursat theorem, theintegral vanishes. Therefore, we have f(k) =

theta(or step) function

{Io

ifk>O, ifk
This is precisely the definition of the theta fnnction (or step function). Thus, we have obtained anintegral representation of that function:

1

00

O(x) = -I . 2Jrl

-00

.

ixt

-e - . dt. t - IE

Now suppose that there are two singular points on the real axis, at XI and xz. Let us avoid XI and Xz by making little semicircles, as before, letting both semicircles be in the UHP (see Figure 10.8). Without writing the integrands, we can represent the contour integral by

The principal value of the integral is naturally defined to be the sum of all integrals having E in their limits. The contribution from the small semicircle Clean be calculated by substituting z - XI = Eeie in the integral:

r

leI

f(z)dz (z - Xl)(Z - Xz) =



f(xI + Eeie)iEe ie de x Eeie(Xl + EeW - Xz) =

.

f(xI) X2'

-In Xl -

with a similar result for CZ. Putting everything together, we get

pl

OO

-00

f(x) d . f(xz)-f(xI) x - l1t (x - Xl)(X - Xz) Xz - XI

'LRes.

= 21i1

If we inclnde the case where both C I and Cz are in the LHP, we get

pl

oo

-00

f(x) dx = (x - XI)(X - Xz)

±.tit f(xz)-f(Xl) + 21!1'LRes, Xz -

Xl

(10.15)

10.3 EVALUATION OF DEFINITE INTEGRALS

289

C2

XI

X2

Figure10.8 One of the four choices of contours for evaluating the principal value of the integral when there are two poles on the real axis.

where the plus sign is for the case where C I and C2 are in the UHP and the minus sign for the case where both are in the LHP. We can also obtain the result for the case where the two singularities coincide by taking the limit Xl --> X2. Then the RHS of the last equation becomes a derivative, and we obtain

P

J

OO

-00

f(x) (x - xo)

2

. ." dx = ±mf'(xo) + 2m L...Res.

10.3.12. Example. An expression encountered in the study of Green's functions or propagators (which we shall discuss later in the book) is eitx dx

00

1

_oox 2 - k 2 '

where k and t are real constants. We want to calculate the principal value of this integral. We use Equation (10.15) and note that for t > 0, we need to close the contour in the UHP, where there are no poles: P

1

00

-00

eitx dx x 2 - k2

---=P

1

00

-00

e itx dx e ikt - e- ikt sinkt =irr --'Jr-(x - k)(x + k) 2k k·

When t < 0, we have to close the contour in the LHP, where again there are no poles:

1

00

P

-00

eitx dx x-2---k-2 = P

1

00

-00

e itx dx e ikt - e- ikt (x - k)(x + k) = -Ln 2k =

sin kt J rk --

·

The two results abovecan be combinedinto a single relation:

1

00

P

eitx dx

sinkltl

-2--2 = - J r - - .

-oox -k

k

11II

290

10. CALCULUS OF RESIOUES

10.4 Problems 10.1. Evaluate each of the following integrals, for all of which C is the circle lel = 3. I

J

(a)

Yc

(d)

i

dz:

sinh z

- 4 - dz.

c z

tanzdz.:

(j)

(m)

2+1

z

c z(z - I)

()i i g

4z-3 z(z _ 2) dz.

i

dz -.-c z2 sin Z'

(b)

(e) (h)

!

i i i

z cos

J

dz

eli

c z(z

(c)

i in) dz: i

cosh z 2

cz

-I-

n

2

(D

(k)

Yc sinh 2z'

(n)

i

(f)

dz:

C (z - I)(z - 2)

cosz

J

Yc

I

(I)

dz:

I-~OSZ dz: z

(") J

dz.

e'dz

J

Yc z(z - n) dz

Yc Z3(Z + 5)'

i:

zdz. cZ

.

10.2. Leth(z) be analytic and have a simple zero at z = zo. and let g(z) be analytic there. Let !(z) = g(z)/ h(z), and show that g(zo) Res[!(zo)] = h'(zo)'

10.3. Find the residne of !(z) = 1/ cos z at each of its poles.

J:'

10.4. Evaluate the integral dx/[(x 2 + l)(x 2 + 4)] by closing the contour (a) in the UHP and (b) in the LHP. 10.5. Evaluate the following integrals, in which a and b are nonzero real constants.

10.4 PROBLEMS

R+ib

o

-R Fi9ure 10.9

291

R+ib

+R

The contour usedin Problem 10.8.

10.6. Evaluate each of the following integrals by turning it into a contour integral around a unit circle.

(2" (a) io 2"

(e)

10o

(e)

10o

(h)

io

(b) {2" d8 io a+cos8 2;< d8

ae

l+sin28' 2" cos2 38

t"

(g)

d8 5+4sin8

ao..

5-4cosZ8 cos 2 3¢ d¢

1- Zacos¢ +

a2

[" cos Z¢ d¢ io 1 - Za cos ¢ + a 2

(i)

10" tan(x + ia) dx

(j)

10" eCos ~ cos(n¢ -

where a > 1.

(d)

10o

(f)

(" d¢ io 1- Za cos ¢ +a 2

(a +bcos2 8 )2

where a, b > O. where a ;f ±1.

where a ;f ± 1. where a ;f ±1. where a E R.

sin ¢) d¢

where n

E

Z.

10.7. Evaluate the integral I = f'''ooeuxdx/(I+eX)forO < a < 1. Hint Choose a closed (long) rectangle that encloses only one of the zeros of the denominator. Show that the contributions of the short sides of the rectangle are zero.

oo x2 10.8. Derive the integration formnla f o e- cos(Zbx)dx = b ;f 0 by integrating the function eFigure 10.9.

z2

V;

e-

b2

where

around the rectangnlar path shown in

10.9. Use the result of Example 10.3.11 to show that 8'(k) = 8(k).

292

10. CALCULUS OF RESIDUES

10.10. Find the principal values of the following iutegrals. (a)

sinxdx

1

(x 2

1

2

00

-00

+ 4) (x

- I)

.

x cosx 5 dx. -oox - x+6 00

(c)

(b) (d)

cosax --dx 3 -00 I +x 00 l-cosx

1 1

00

x

-00

2

where a "': O.

dx.

10.11. Evaluate the following integrals.

rOO xl _

(a) Jo (c)

1 1

00

2

(Sinax)

x2 + b2 - x - dx. sin ax

o x(x 2 + b2)2 00

(e)

b

o

dx

.

8in2 x dx

x2

'

(b)

roo

sin ax d x(x 2 + b2) x. 00 cos 2ax - cos 2bx d 2 x. o x sin 3 x dx · x3

Jo (d)

(f)

1roo 10

Additional Reading I. Dennery, P. and Krzywicki, A. Mathematics for Physicists, Harper and Row, 1967. Includes a detailed discussion of complex analysis encompassing applications of conformal mappings and the residue theorem to physical problems. 2. Mathews, J. and Walker, R. Mathematical Methods of Physics, 2nd ed., Benjamin, 1970. A "practical" guide to mathematical physics, including a long discussion of "how to evaluate integrals" and the use of the residue theorem.

11

_

Complex Analysis: Advanced Topics

The subject of complex analysis is an extremely rich and powerful area of mathematics. We have already seeu some of this richuess and power io the previous chapter. This chapter concludes our discussion of complex analysis by iotroducing some other topics with varyiog degrees of importance.

11.1 Meromorphic Functions meromorphic functions

Complex functions that have only simple poles as their siogularities are numerous in applications and are called meromorpbic functions. In this section, we derive an important result for such functions. Assume that! (z) has simple poles at{z j }f~l' where N could be iofinity. Then, if z

01 Z j

~ 2Jfl

for all j, the residue theorem yields! (

!CI;) dl; = !CZ)+ tRes eCI;)) , Z j=! I; - Z ~=Zj

len I; -

where Cn is a circle contaioiog the first n poles, and it is assumed that the poles are arranged in order of iocreasiog absolute values. Since the poles of ! are assumed to be simple, we have ! ~) ) Res - I; - Z ~=Zj

(

. = Inn

~""Zj

CI; -

1

!~)

Zj)--. I; - Z

= -1Zj -

= --Res[!Cf;)]~~Zj es Zj - Z

INote that theresidue of f(~)/(~ - z) at';

=

z is simply f(z),

. lim [CI; -

Z ~""Zj

r'

_J_,

Zj -

Z

zj)!CI;)]

294

11. COMPLEX ANALYSIS: ADVANCED TOPICS

where rj is, by definition, the residue of f(~) at ~ = ij. Substituting in the preceding equation gives

f(z) =

~ 2:",

(

f(~) d~ - t~.

Je, ~ - Z

j~l Zj - Z

Taking the difference between this and the same equation evaluated at Z = 0 (assumed to be none of the poles),Z we can write

f(z) - f(O) =

-f--, len( H~(~»7Cl

Mittag-Leffler expansion

d~ +

Z

t

)=1

rs (_1_ + Z - Zj

2.) . Zj

If If(~)1 approaches a finite value as I~I --.. 00, the integral vanishes for an infinite circle (which includes all poles now), and we obtain what is called the MittagLeffler expansion of the meromorphic fuoction f: f(z) = f(O)

(_1_ + 2.).

+ trj )=1

Z - Zj

(11.1)

Zj

Now we let g be an entire function with simple zeros. We claim that (a) (dg/dz)/g(z) is a meromorphic function that is bounded for all values of z, and (b) its residues are all unity. To see this, note that g is of the fomr'

g(z) = (z - ZI)(Z - zz)··· (z - zN)f(z), where Zl, ... , ZN are all the zeros of g, and f is an analytic function that does not vanish anywhere in the complex plane. It is now easy to see that

g'(z) N I z - z· g (z ) =

L

J=l

j

+

f'(z) f(z) .

This expression has both properties (a) and (b) mentioned above. Furthermore, the last term is an entire function that is bounded for all C. Therefore, it must be a constant by Proposition 9.5.5. This derivation also verifies Equation (11.1), which in the case at hand can be written as

d -d Ing(z) z

s'(z)) = -d d N = -( Ing(O)+ L g z Z j~l

(I -- + -I) , Z - Zj

Zj

whose solution is readily found to be

g(Z) = g(O)eCZ

fI (I - ~)

)=1

and it is assumed that Zj

e Z/ Zj

z)

where

c=

(dg/dz)lz~o

g(O)

(11.2)

i= 0 for all j,

~This is not a restrictive assumption because we can always move our coordinate system so that the origin avoids all poles. One can ''prove'' this by factoring the simple zeros one by one, writing g(z) = (z - Zl)!t (z) and noting that g(Z2) = 0, with Z2

i= at. implies that il (z)

= (z - Zz) h (z), etc.

11.2 MULTIVALUED FUNCTIONS

(a)

295

(b)

Figure 11.1 (a) Theangle eO changes by 2rr as ZQ makes a complete circuit around C. (b) The angle80 returns to its original valuewhen zQ completes the circuit.

11.2 Multivalued Functions The arbitrariness, up to a multiple of2"., of the angle 0 = arg(z) inz = re ie leads to fuuctions that can take different values at the same point. Consider, for example, the function !(z) = .,ji.. Writing z in polar coordinates, we obtain !(z) = f tr, 0) = (re ie)t/ 2 = ..;reie / 2. This shows that for the same z = (r,O) = (r,O + 2".), we get two different values, f tr, 0) and f'(r, 0 + 2".) = - !(r, 0). This may be disturbing at first. After all, the definition of a function (mapping) ensures that for any point in the domain a unique image is obtained. Here two different images are obtained for the same z. Riemann found a cure for this complex "double vision" by introducing what is now called Riemann sheets. We will discuss these briefly below, but first let us take a closer look at a prototype of multivalued functions. Consider the natural log function, In z. For z = re ie this is defined as In z = In r +iO = In [z] +i arg(z) where arg(z) is defined only to within a multiple of 2".; that is, arg(z) = 0 + Znit , for n = 0, ±l, ±2, .... We can see the peculiar nature of the logarithmic function by considering a closed path around the point z = 0, as shown in Figure 11.1(a). Starting at Zo, we move counterclockwise, noticing the constant increase in the angle (}o, until we reach the initial point in the z-plane. However, the angle is then 00 + 2".. Thus, the process of moving around the origin has changed the value of the log function by 2".i. Thus, (lnzo)final - (lnzo)mitiol = Zati. Note that in this process zo does not

change, because

branch point

11.2.1. Definition. A branch point ofafunction f

: iC -> iC is a complex number

296

11. COMPLEX ANALYSIS: ADVANCED TOPICS

zo with the property that f(ro, 110) oft f(ro, 1I0 + 211") for any closed curve C encircling Z00 Here (ro, 110) are the polar coordinates ofzo.

Victor-Alexandre Pniseox (1820-1883) was the first to take up the subject of multivalued functions. In 1850 Puiseux published a celebrated paper on complex algebraic functions givenby f (u, z) = 0, f apolynomial in u andz. Hefirst made clearthedistinction between poles andbranchpointsthat Cauchy hadbarely perceived, andintroduced thenotion of an essential singular point,to which Weierstrass independently hadcalledattention. Though Cauchy, in the 1846paper, didconsider thevariation of simple multivalued functions along paths that enclosed branch points, Puiseux clarified this subject too. Puiseux also showed that thedevelopment of afunction of z about a branch point z = a must involve fractional powers of z - a. He thenimproved on Cauchy's theorem on the expansion of a function in a Maclaurin series. By his significant investigations of manyvalued functions and their branch points in the complex plane, andby his initial work on integrals of suchfunctions, Puiseux brought Cauchy's pioneering work in function theory to the end of what might be called the first stage. The difficnlties in the thenry of multiplevalued functions and integrals of suchfunctions werestillto be overcome. Cauchy didwrite other papers on theintegrals of multiplevalued functions in whichhe attempted to follow upon Puiseux's work; andthough he introduced thenotion of branch cuts (!ignes d~arrh), be was still confused about the distinction between poles and branch points. Thissubject of algebraic functions andtheir integrals was to be pursued by Riemann. Puiseux was a keenmountaineer andwas thefirst to scalethe Alpine peak that is now named after him.

Thus, Z = 0 is a branch point of the logarithmic function. Studying the behavior of 1n(I/z) = -Inz around z = 0 will reveal that the point "at infinity" is also a branch point of In z. Figure 11.1(b) shows that any other point of the complex )lIane, such as z', carmot be a branch point because 110 does not change when C' is traversed completely.

11.2.1 Riemann Surfaces

branch cut or simply "cut"

The idea of a Riemarm surface begins with the removal of all points that lie on the line (or any other curve) joining two branch points. For In z this means the removal of all points lying on a curve that starts at Z = 0 and extends all the way to infinity. Such a curve is called a branch cut, or simply a cut. Let us concentrate on In z and take the cut to be along the negative half of the real axis. Let us also define the functions

fn (z) =

I« (r, II)

= Inr + i(1I + 2n1l")

for -

1C

<

e<

1C; r > 0; n = 0, ±1, ... ,

11.2 MULTIVALUED FUNCTIONS

297

so in (z) takes on the same values for -:rr: < e < n that In z takes in the range (2n - I):rr: < e < (2n + I):rr:. We have replaced the multivalued logarithmic function by a series of different functions that are analytic in the cut z-plane. This process of cutting the z-plane and then defining a sequence of functions eliminates the contradiction caused by the existence of branch points, since we are no longer allowed to completely encircle a branch point. A complete circulation involves crossiog the cut, which, in tum, violates the domain of definition of in (z). We have made good progress. We have replaced the (nonanalytic) multivalued function In Z with a series of analytic (in their domain of definition) functions in (z). However, there is a problem left: in(z) has a discontinuity at the cut. In fact, just above the cut in(r,:rr: - E) = Inr +i(:rr: - E +2n:rr:) with s c- 0, and just below it in(r, -:rr: + E) = Inr + i(-:rr: + E + 2n:rr:), so that lim [in(r,:rr: - E) - in(r, -:rr: + E)l = 2:rr:i. ' ....0

Riemann surfaces and sheets

To cure this we make the observation that the value of in (z) just above the cut is the same as the value of in+1 (z) just below the cut. This suggests the following geometrical construction, due to Riemarm: Superpose an infinite series of cut complex planes one on top of the other, each plane corresponding to a different value of n. The adjacent planes are connected along the cut such that the upper lip of the cut in the (n - I)th plane is connected to the lower lip of the cut in the nth plane. All planes contain the two branch points. That is, the branch points appear as "hinges" at which all the planes are joined. With this geometrical construction, if we cross the cut, we end up on a different plane adjacent to the previous one (Figure 11.2). The geometric surface thus constructed is called a Riemann surface; each plane is called a Riemann sheet and is denoted by R], for j = 0, ±I, ±2, .... A single-valued function defined on a Riemarm sheet is called a branch of the original multivalued function. We have achieved the following: From a multivalued function we have constructed a sequence of single-valued functions, each defined in a single complex plane; from this sequence of functions we have constructed a single complex function defined on a single Riemarm surface. Thus, the logatithmic function is analytic throughout the Riemarm surface except at the branch points, which are simply the function's singular points. It is now easy to see the geometrical significance of branch points. A complete cycle around a branch point takes us to another Riemarm sheet, where the function takes on a different form. On the other hand, a complete cycle around an ordinary pointeither nevercrosses the cut, or if it does, it will cross it backto the original sheet. Let us now briefly consider two of the more common multivalued functions and their Riemarm surfaces. 11.2.2. Example. THE FUNCTION f(z) = z'!" The only branch points for the functioo f(z) = zl/n are z

= 0 and the point at infinity.

298

11. COMPLEX ANALYSIS: ADVANCED TOPICS

Figure 11.2 - A few sheets of the Riemann surface of the logarithmic function. The path C encircling the origin 0 ends up on the lower sheet.

Defioing !k(z) sa r 1/ nei (9+2k" jn) for k = 0, 1, ... , n -1 and 0 <

e

< 2". and following the same procedure as for the logarithmic function, we see that there must be n Riemann sheets, labeled Ro. RI •... , Rn-l. in the Riemann surface. The lower edge of Rn-l is pasted to the upper edge of Ro along the cut, which is taken to be along the positive real axis. The Riemann surface for n = 2 is shown in Figure 11.3. It is c1ear that for any nonintegervalue of a the function fez) = zOl. has a branch point at z = 0 and another at the point at infinity. For irrational a the number of Riemann sheets is infinite. II

11.2.3. Example. l'HEFUNCTION f(z) = (z2 _1)1/2 The brancb points for the function f(z) = (z2 - 1)1/2 are at zt = +1 and Z2 = -1 (see Figure 11.4).Writing z - 1 = rt ie, and z + 1 = r2eie,. we bave f(z) = (Yjeie1)1/2 (r2eie,)I/2 = .jYj r 2ei (e1+e, )/ 2. The cut is along the real axis from z = -1 to z = + 1. There are two Riemann sheets in the Riemann surface. Clearly, only cycles of 2rr involving one branch point will cross the cut and therefore end up on a different sheet. Any closed curve that has both zr and Z2 as II interior points will remain entirely on the original sheet.

evaluation of integrals involving Guts

The notion of branch cnts can be used to evalnate certain integrals that do not fit into the three categories discussed in Chapter 10. The basic idea is to circumvent the cut by constructing a contour that is infinitesimally close to the cut and circles around branch points.

11.2 MULTIVALUED FUNCTIONS

Figure 11.3

299

The Riemann surface for fez) = zl/2.

z

Figure 11.4 The cut for the fuoction fez) = (z2 - 1)1/2 is from ZI to Z2. Paths that circle only one of the points cross the cut and end up on the other sheet.

10

00 adx/(x2 11.2.4. Example. To evaluate the integrall = x + I) for lal < I, consider adz/(z2 the complex integral t' = :Fe z + 1) where C is as shown in Figure 11.5 and the cut is taken along the positive real axis. To evaluate the contribution from C R and Cr, we let p stand for either T or R. Then we have

It is clear that since lal < 1, I p -+ 0 as p -)- 0 or p -+ 00.

300

11. COMPLEX ANALYSIS: ADVANCED TOPICS

Figure 11.5

Thecontour fortheevaluation ofthe integrals of Examples 11.2.4and11.2.5.

The contributions from L 1 and £2 do not cancel one another because the value of the functionchanges above andbelow the cut. To evaluate these two integrals we have to choose a branch of the function. Let us choose that branch on which ZCl = IzlfteiaB for o < () < 2Jt. Along L1, (J ~ 0 or zet = x", and along L2. e ~ 2rr or zet = (xe 27ti)a'. Thus,

looo -2-x'd x + 1°

i

z· --4Z= 0 C z2 + I

x

+I

= (1 - e21ri Cl. )

00

ct 2rria x e.

(xe 2n')2 + I

dx

tX) ~ dx, 2

10

x

(11.3)

+I

TheLHSof this equation canbe obtained using theresidue theorem. There are two simple poles, at z = +i and z = -i with residues Res[f(i)] = (ein /2). /2i and Res[f(-i)] _(ei3n/2). /2i. Thus,

1i



- - dz = 2rci

cz2 + 1

(e-i•n/2 e i3• n/ 2) - .- - - - . 21 21

= rt(e w. rrj 2 _

=

.

eI3car/2).

Combiniog this with Equation (11.3), we obtaio

[00 ~ dx = 7r(eicmj2 -

io xl + 1

<3urr/2) = !!:. sec a7r . 1 - e~bna 2 2

If we hadchosena different branch of thefunction, boththeLHSandtheRHSof Equation (11.3) wonld have been different, but the fina1 result would stiD have been the same. III

11.2 MULTIVALUED FUNCTIDNS

301

Ce

The contour for the evaluation of the integral of Example 11.2.6.

Figure 11.6

11.2.5. Example. Here is another integral involving a branch cut: 1=

roo

10

x-a dx x+ I

for 0 < a < 1.

To evaluate this integral we use the zeroth branch of the function and the contour of the previous example (Figure 11.5). Thus, writing z = peW. we have

2"iRes[f(-I)]

=

J:

z-a dz

Ic z + +

=

I

[00

10

p-a dp p+ I

0 (pe 2i,,)-a. 2mdp+

1 00

.

pe Zm

+1

e

+ J:

z-a dz

+ I a c --dz. Cr Z + 1

i

fCR Z

(11.4)

The contributions from both circles vanish by the same argument used in the previous example. On the other hand, Res[f( -I)] = (_I)-a. For the branch we are using, -I = e i". Thus, Res[f( -I)] e-ia". The RHS of Equation (11.4) yields

=

00

10o

-a

_P_ dp - e-2in a

10

p+1

00

0

-a ..!!...:.....dp

= (1 _ e- 2i1ta)I.

p+1

It follows from (11.4) that (1 - e- 2iJra)I = 21rie-i 1r Q , or 00

10o

x-a x+l

1C

for 0 < a < 1.

dx - sinarr

III

oo 11.2.6. Example. Let us evaluate I = o Inx dx/(x 2 +a 2) with a > O. We choose the zeroth branch of the logarithmic function, in which - l i < 9 < Jr, and use the contour of Figure 11.6. For Ll' Z = pe i1r (note that p > 0), and for £2, Z = p. Thus, we have

f

2"i Res[f(ia)]

=

i

C

+

ln z

-2--2 dz

z +a

1

00

E

=

1<

ln z e" dp + /,C" ~ dz z +a

In(pe;").

.

2

2

(pelJt ) +a lup /, Inz -2--2 dp + ~ dz. p +a CR z +a 00

(11.5)

302

11. COMPLEX ANALYSIS: ADVANCED TOPICS

where z = ia is the only singularity-a simple pole-in the UHP. Now we note that

l

E

In(pei:Jr).l1C . e dp = ( pe l'lr) 2 +a 2

00

1

00

to

lop + itt dp= p2 +a 2

1

00

E

lop . ---dp+zJr p2 +a 2

1

00

to

dp ---. p2 +a 2

Thecontributions from the circlestendto zero. On the other hand, Res[f(ia)] = lim (z - ia)

z-via

. In z

(z - w)(z

In(ia) 1 ( Ina +!''') . = .- = -.2za 'Iia 2

.

+ za)

Substituting thelast two results in Equation (11.5), we obtain

;llP+a1 dp +i:7r1

00

?:'(lna+i::')=21°O 2

a

E

P

E

p

zd

p

+a

2"

It can also easilybe shownthat ooo dp/(p2 + a2) = ,,/(2a). Thus, in the limit E -+ 0, we

I

= !!..-In a. The sign of a is irrelevant because it appears as a square in the integral. 2a Thus, we can write 00 Inx n a i' O. -2--2 dx = -21Ilnla l,

get I

1 D

x +a

a

n.3 Analytic Continuation Analytic functions have certain unique properties, some of which we have already noted. For instance, the Cauchy integral formula gives the value of an analytic function inside a simple closed contour once its value on the contour is known. We have also seen that we can deform the contours of integration as long as we do not encounter any singnlarities of the function. Combining these two properties and assuming that f : C -+ C is analytic within a region See, we can ask the following question: Is it possible to extend f beyond S1 We shall see in this section that the answer is yes in many cases of interest." First consider equal on a piece, equal all over

1I.3.1. Theorem. Let fl' fz : C -+ C be analytic in a region S. If fl = I: in a neighborhood ofa point z E S, or for a segment ofa curve in S, then I, = fz for

all z

E

S.

Proof Let g = f1 - fz, and U = {z E S I g(z) = OJ. Then U is a subset of S that includes the neighborhood of z (or the line segment) in which fl = h. If U is the entire region S, we are done. Otherwise, U has a boundary beyond which g(z) oF O. Since all points within the boundary satisfy g(z) = 0, and since g is continuous (more than that, it is analytic) on S, g must vanish also on the boundary. But the bouodary points are not isolated: Any small circle around anyone of them includes points of U as well as points outside U. Thus, g must vanish on a neighborhood of any boundary point, implying that g vanishes for some points outside U. This 0 contradicts our assumption. Thus, U must include the entire region S. 4Provided that S is not discrete (countable). (See [Lang 85, p. 91].)

11.3 ANALYTIC CONTINUATION

303

A consequence of this theorem is the following corollary. 11.3.2. Corollary. The behavior 01afunction that is analytic in a region S c C is completely determined by its behavior in a (small) neighborhood 01an arbitrary point in that region. analytic continuation

This process of detetmining the behavior of an analytic function outside the region in which it was originally defined is called analytic continuation. Although there are infinitely many ways of analytically continuing beyond regions of definition, the values of all functions obtained as a result of diverse continuations are the sarne at any given point. This follows from Theorem 11.3.1. Let It, Iz : C -+ C be analytic in regions 8t and 82, respectively. Suppose that II and fz have different functional forms in their respective regions of analyticity. If there is an overlap between 81 and 82 and if II = fz within that overlap, then the (unique) analytic continuation of II into 82 must be h. and vice versa. In fact, we may regard !J and h as a single function I : iC -+ C such that I(z) =

{!J(Z) I: (z)

when Z E 81, when Z E 82.

Clearly, I is analytic for the combined region 8 = 81 U 82. We then say that and fz are analytic continuations of one another.

II

11.3.3. Example. Let us consider the function 11 (z) = L~o t", which is analytic for lel < 1. We have seen that it converges to 1(1 - z) for [z] < 1. Thus, we have II (z) = 1((1 - z) when [z] < I, and!J is not definedfor Izi > 1. Now let us consider a second function, h(z) = L;:'=o (~)n+1 (z + i)n, which converges for [z + ~I < ~. To see what it converges to, we note that h(z) =

~ L~o [~(z + i)]", Thus, h(z) =

~ I-~(z+~)

I =-

when Iz+il <

I-z



We observethat although II (z) and h (z) have different series representations in the two overlapping regions (seeFigure 11.7), they representthe samefunction, I(z) = 1(1- z). Wecantherefore write

I

fJ (z) when Izl < I, - { h(z) when lz + ~I <

(z) -

i,

and I: and h are analytic continuations of ooe another. In fact, I(z) = 1(1 - z) is the analytic continuation of both fJ and Iz for all of iC except z = 1. Figure 11.7 sbows Si, the region of definition of

Ii. for i

= 1, 2.

III

11.3.4. Example. The function fJ (z) = 1000 e- zt dt exists only if Re(z) > 0, in which case II (z) = I(z. Its regionof definitionSI is shownin Figure H.S and is simplythe right half-plane.

304

11. COMPLEX ANALYSIS: ADVANCED TOPICS

Figure 11.7 Thefunction defined in thesmaller circleis continued analytically intothe larger circle.

Figure 11.8

Thefunctions

!l and I: are analytic continuations of each other: !l ana12 analytically continues !J into the

lytically continues 12 into the right half-plane, and senticirc1ein the left half-plane.

Now we define 12 by a geometric series: h(z) = i L:~O[(z+i)fi]n where Iz+il Thisseries converges, within its circleof convergence 82, to . l

I I =-. l-(z+iJfi z

<

I.

11.3 ANALYTIC CONTINUATION

(a)

305

(b)

(a) Regions 81 and 82 separated by the bonndary B and the contnnr C. (b) Thecontour C splitsup into Cl andC2-

Figure 11.9

Thus, we have ~={ft(Z) z h(z)

when z e Sj , when Z E 82.

Thetwofunctions are -analytic continuations of oneanother, andf (z) = 1/ z is theanalytic continoation of both II and 12 for aU z E C except z = O. iii

11.3.1 The Schwarz Reflection Principle A result that is useful in some physical applications is referred to as a dispersion relation. To derive such a relation we need to know the behavior of analytic functions on either side of the real axis. This is found using the Schwarz reflection principle, for which we need the following result. 11.3.5. Proposition. Let Ii be analytic throughout 8i, where i = I, 2. Let B be the boundary between 81 and 82 (Figure 11.9) and assume that It and Iz are continuous on B and coincide there. Thenthe twofunctions are analytic continuations ofone another and together they define a (unique)function

I(z) = {fl (z) h(z)

when z E 81 U B, when z E 82 U B,

which is analytic throughaut the entire region 81 U 82 U B.

306

11. COMPLEX ANALYSIS: ADVANCED TOPICS

Proof The proof consists in showing that the function integrates to zero along any closed curve in Sl U S2 U B. Once this is done, one can use Morera's theorem to conclude analyticity. The case when the closed curve is entirely in either S; or S2 is trivial. When the curve is partially in Sj and partially in S2 the proof becomes only slightly more complicated, because one has to split up the contour C into Cl and C2 of Figure 11.9(b). The details are left as an exercise. D Schwarz reflection principle

11.3.6. Theorem. (Schwarz reflection principle) Let f be a function that is analytic in a region S that has a segment ofthe real axis as part ofits boundary B. If f(z) is real whenever z is real, then the analytic continuation g of f into S* (the mirror image of S with respect to the real axis) exists and is given by

g(z) = f*(z*)

where z E S*.

Proof First, we show that g is analytic in S*. Let f(z) sa u(x, y)

+ iv(x, y),

g(z) es U(x, y)

+ iV(x, y).

Then f(z*) = fix, -y) = u(x, -y) + iv(x, -y) and g(z) = f*(z*) imply that U(x, y) = u(x, -y) and V(x, y) = -v(x, -y). Therefore,

au au av av av = = =---=ax ax ay a(-y) ay' au au av av = - ay = ax = -""h'

ay

These are the Cauchy-Riemann conditions for g(z). Thus, g is analytic. Next, we note that fix, 0) = g(x, 0), implying that f and g agree on the real axis. Proposition 11.3.5 then implies that f and g are analytic continuations of one another. D It follows from this theorem that there exists an analytic function h such that

h(z) = {f(Z) g(z)

when z E S, when z E S*.

We note that h(z*) = g(z*) = f*(z) = h*(z).

11.3.2 DispersionRelations Let f he analytic throughout the complex plane except at a cut along the real axis extending from xo to infinity. For a point z not on the x-axis, the Cauchy integral f(~) d~ I(~ - z) where C is the contour shown formula gives f(z) = (2ni}-1 in Figure 11.10. We assume that f drops to zero fast enough that the contribution

Ie

11.3 ANALYTIC CONTINUATION

307

o

The contour used for dispersion relations.

Figure 11.10

from the large circle tends to zero. The reader may show that the contribution from the small half-circle around XQ also vanishes. Then

f(z) = -1. 2rn

[l""+i' f(~)

-d~

xo+i'e ~ - Z

-

roo

l""-i' f(~)z xo-ie

roo

= ~ [ f(x + i~) dx _ 2nl JXD x - Z + IE Jxo

-d~ ] ~ -

f(x - if) dX]. X - Z - IE

Since z is not on the real axis, we can ignore the i E tenus in the denominators, so that f(z) = (2rri)-1 Jx':[f(x + if) - f(x - if)]dxj(x - z). The Schwarz reflection principle in the form J*(z) = f(z*) can now be used to yield f(x dispersion relation

+ if) -

f(x - if)

= f(x + if) -

J*(x

+ if) = 2i Im[f(x + if)].

The final result is

1 f(z) = tt

1"" xo

Im[f(x

+ if)] . dx.

x-z

This is one form of a dispersion relation. It expresses the value of a function at any point of the cut complex plane in terms of an integral of the imaginary part of the function on the upper edge of the cut. When there are no residues in the UHP, we can obtain other forms ofdispersion relations by equating the real and imaginary parts of Equation (l0.11). The result

308

11. COMPLEX ANALYSIS: ADVANCED TOPICS

is

1 =F~P 1

00

1 Re[f(xo)] = ±-P Jr

Im[f(x)] X

-00

-xo

dx,

00

Im[f(xo)] =

1T:

Hilbert transform

Re[f(x)] dx,

(11.6)

x-xo

-00

where the upper (lower) sign corresponds to placing the small semicircle around xo in the UHP (LHP). The real and imaginary parts of f, as related by Equation (11.6), are sometimes sometiroes said to be the Hilbert transform of one another. In some applications the imaginary part of f is an odd function ofits argument. Then the first equation in (11.6) can be wtitten as 2 Re[f(xo)] = ±-P x

10

00

0

x Im[f(x)] 2 dx. x 2-xO

To artive at dispersion relations, the following condition must hold: lim Rlf(Reie)1 = 0,

R->oo

. where R is the radius of the large semicircle in the UHP (or LHP). If f does not satisfy this prerequisite, it is still possible to obtain a dispersion relation called dispersion relation a dispersion relation with one subtraction. This can be done by introducing with onesubtraction an extra factor of x in the denominator of the integrand. We start with Equation (l 0.15), confining ourselves to the UHP and assuming that there are no poles there, so that the sum over residues is dropped:

f(X2) - f(XI) = ~P X2 - Xl in

1

00

-00

f(x) dx, (x - XI)(X - X2)

The reader may check that by equating the real and imaginary parts on both sides, letting Xl = 0 and X2 = xo, and changing X to -x in the first half of the interval of integration, we obtain

:.. .Re.:. o[::.. f.:c.(x-",o)",] = Re[f(O)] Xo xo

+ ~ [p n

roo Im[f(-x)] dx + P roo Im[f(x)] dX].

10

x(x

10

+ xo)

x(x - xo)

For the case where Im[f( -x)] = -lm[f(x)], this equation yields

Re[f(xo)] = Re[f(O)]

+ 2x~ P n

optical theorem

roo

10

Im~f(X)1

x(x -xo)

dx.

(11.7)

11.3.7. Example. In optics. it has been shown that the imaginary part of the forwardscatteringlight amplitudewith freqnency 00 is related, by the so-called optical theorem, to the total cross section for the absorptionof light of that frequency: 00

!m[f(oo)] = -<1lot(oo).

411"

11.4 THE GAMMA AND BETA FUNCTIONS

Kramers-Kronig relation

309

Substituting this in Equation(t 1.7)yields Re[f(wo)] = Re[f(O)]

,,} p looo + ----%: 2:rr

0

O"totCw) 2 2 da.

(11.8)

(V-coO

Thus, the real part of the (coherent) forward scatteringof light, that is, the real part of the indexcfrefraaion. canbe computedfromEquation(11.8)byeithermeasuring orcalculating O'tot(w), the simpler quantity describing the absorption of light in the medium. Equation (11.8) is the original Kramers-Kronigrelation. II

11.4 The Gamma and Beta Functions

gamma function defined

We have already encountered the gamma function. In this section, we derive some useful relations involving the gamma function and the closely related beta function. The gamma function is a generalization of the factorial function-which is defined only for positive integers-to the system of complex numbers, By differentiating oo the integral 1(01) '" fo e-aldt = 1/01 with respect to 01 repeatedly and setting 01 = I at the end, we get f ooo t ne-I dt = nL This fact motivates the generalization r(z) '" faoo t,-Ie-Idt

for Refz) > 0,

(11.9)

where I' is called the gamma (or factorial) function. It is also called Euler's integral ofthe second kind. It is clear from its definition that r(n

+ I)

= n!

(11.10)

if n is a positive integer. The restriction Re(z) > 0 assnres the convergence of the integral. An immediate consequence of Equation (11.9) is obtained by integrating it by parts:

I'(z + I) = zF'(z).

(11.11)

This also leads to Equation (11.1 0) by iteration. Another consequence is the analyticity of I'(z), Differentiating Equation (11.11) with respectto z, we obtain

dr(z + 1) r() dr(z) dz = z +zdZ" Thus, dr(z)/dz exists and is finite if and only if dr(z + I)/dz is finite (recall that z oF 0). The procednre of showing the latter is outlined in Problem 11.16. Therefore, F'(z) is auaiytic whenever I'(z + 1) is. To see the singularities of I'(z), we note that

I'{z

+ n) =

z(z

+ l)(z + 2) ... (z + n -

I)r(z),

310

11. COMPLEX ANALYSIS: ADVANCED TOPICS

or F'(z} = z(z

f(z + n) + l)(z + 2) ... (z + n -

(11.12)

1)

The numerator is analytic as long as Re(z + n) > 0, or Retz) > r--Fl, Thus, for Refz) > -n, the singularities of T'(z) arethepolesatz = 0, -I, -2, ... , -n+ I. Since n is arbitrary, we conclude that

11.4.1. Box. I'(z) is analytic at all Z E iC except at z = 0, -1, -2, ... , where I'(z) has simple poles.

A useful result is obtained by setting z =

! in Equation (11.9): (11.13)

This can be obtained by making the substitntion u = .,fi in the integral. We can derive an expression for the logarithmic derivative of the gamma function that involves an infinite series. To do so, we use Equation (11.2) noting that 1/ I'(z + I) is an entire function with simple zeros at {-k}~I' Equation (11.2) gives

where y is a constant to be determined. Using Equation (11.11), we obtain

Z)

f1

-1- = zeyZ 00 ( 1 + - e-z/ k • I'(z) k~1 k

Euler-Mascheroni constant

To determine y, let z = 1 in Equation (11.14) and evaluate the resulting product numerically. The result is y = 0.57721566 ... , the so-called Euler-Mascheroni constant. Differentiating the logarithm of both sides of Equation (11.14), we obtain d -in[f(z)] = - -1 - y

dz beta function defined

(11.14)

Z

+ I: (1- - - 1 ) . 00

k=l

k

z +k

(11.15)

Other properties ofthe ganuna function are derivable from the results presented here. Those derivations are left as problems. The beta function, or Euler's integral of the first kind, is defined for complex numbers a and b as follows:

B(a, b) es

fal t a- 1(1 -

t)b- 1dt

where Re(a), Re(b) > O.

(11.16)

311

11.4 THE GAMMA AND BETA FUNCTIONS

By changing I to 1/1, we can also write

1

00

R(a, b) sa

I-a-b(t - l)b-1dl.

(11.17)

Since 0 ::: I ::: I in Equation (11.16), we can define R(a, b) = 2

["/2

J

e by I =

sin 2 e. This gives

sin2a- 1 e cos2b- 1e de.

(11.18)

o

This relation can be used to establish a connection between the gamma and beta functions. We note that

gamma function and beta function are related

where in the last step we changed the variable to x = .;t. Multiply I'(c) by ['(b) and express the resulting double integral in terms of polar coordinates to obtain ['(a)[,(b) = ['(a + b)R(a, b), or R(a, b) = Bib, a) =

['(a)[,(b) ['(a

+ b)

.

(11.19)

Let us now establish the following useful relation: 11:

[,(z)[,(1 - z) = -.-.

(11.20)

SIDrr:Z

With a = z and b = 1- z, and using u = tan s, Equations (11.18) and (11.19) give

[,(z)[,(1 - z)

= R(z, I -

(00

z)

=2J

o

u2z~1

u2

+ I du

for 0 < Re(z) < I.

Using the result obtained in Exaruple 11.2.4, we immediately get Equation (11.20), valid for 0 < Re(z) < I. By analytic continuation we then geueralize Equation (11.20) to values of z for which both sides are analytic. 11.4.2. Example. As an illustration of the use of Equation(tt .20),tet us showthat I' (z) canalso be written as 1

1

r(z) = 2:n:i

[

e'

(11.21)

lc Ii dt,

whereC is the contourshownin Fignre11.11.FromEqoations (11.9)and (tt .20)it follows that

10

10

1 sin xz sin rrz 00 e i1rz - e- i 7l'Z 00 e- r = - - r ( l - z) = - «" r-'dr = dr. rez):n: :n: 0 2:n:i 0 rt

312

11. COMPLEX ANALYSIS: ADVANCED TOPICS

Im(t)

c

Figure 11.11

The contour C used in evaluating thereciprocal gamma function.

The contour integral of Equation (11.21) can be evaluated by noting that above the real axis, t = rein = -r, below it t = re-i:n: = -r. and, as the reader may check, that the contribution from the smallcircle atthe origin is zero; so

l

r r loco -.-(-dr) e+ 1° e-. (-dr) 0 (re Y (re-m)Z r . loco eloco e-r dr. . = dr +

et -dt=

C tZ

l1T

_e- l1rz

o

00

_

rZ

el.7rZ

-

0

rZ

Comparison with thelastequation aboveyields the desired result.

III

Another useful relation can be obtained by combining Equations (11.11) and (11.20): r(z)r(1 - z) = r(z)( -z)r(-z) = rr/ sin zrz. Thus, r(z)r(-z) =

ic

z sin zrz

(11.22)

Once we know rex) for positive values of real x, we can use Equation (11.22) to find rex) for x < O. Thus, for instance, r(!) =.,fii gives r(-!) = -2.,fii. Equation (11.22) also shows that the gamma function has simple poles wherever z is a negative integer.

11.5 Method of Steepest Descent It is shown in statistical mechanics ([Hill 87, pp. 150-152]) that the partition function, which generates all the thermodynamical quantities, can be written as a contour integral. Debye found a very elegant technique of approximating this contour integral, which we investigate in this section. Consider the integral I (a)

==

fc

eU!(z) g(z)

dz

(11.23)

where lal is large and f and g are analytic in some region of C containing the contour C. Since this integral occurs frequently in physical applications, it would

11.5 METHOD OF STEEPEST DESCENT

313

be helpful if we could find a general approximation for it that is applicable for all f and g. The fact that lal is large will be of great help. By redefining f(z), if necessary, we can assume that a = lale' arg(a) is real and positive [absorb e' argte) into the function f(z) if need be]. The exponent of the integrand can be written as af(z) = au(x, y)

+ iav(x, y).

Since a is large and positive, we expect the exponential to be the largest at the maximum of u(x, y). Thus, if we deform the contour so that it passes through a point zo at which u (x, y) is maximum, the contribution to the integral may come mostly from the neighborhood of zoo This opens up the possibility of expanding the exponent about zo and keeping the lowest terms in the expansion, which is what we are after. There is one catch, however. Because of the largeness of a, the imaginary part of af in the exponent will oscillate violently as v(x, y) changes even by a small amount. This oscillation can make the contribution of the real part of f(zo) negligibly small and render the whole procedure useless. Thus, we want to tame the variation of exp[iv(x, y)] by making v(x, y) vary as slowly as possible. A necessary condition is for the derivative of v to vanish at zoo This and the fact that the real part is to have a maximum at zo lead to au ax

. av

+ Z ax

dfl = dz zo =

(11.24)

o.

However, we do not stop here but demand that the imaginary part of f be constant along the deformed contour: Im[f(z)] Im[f(zo)] or v(x, y) v(xo, Yo). Equation (11.24) and the Cauchy-Riemann conditions imply that aujax = 0= aujay atzo. Thus, it might appear that zo is a maximum (or minimum) of the surface described by the function u(x, y). This is not true: For the surface to have a maximum (minimum), both second derivatives, a2ujax2 and a2uja y 2, must be negative (positive). But that is impossible because u(x, y) is harmonic-the sum of these two derivatives is zero. Recall that a point at which the derivatives vanish but that is neither a maximum nor a minimum is called a saddle point. That is why the procedure described below is sometimes called the saddle point approximation.

=

=

We are interested in values of z close to zo. So let us expand f(z) in a Taylor series about zo, use Equation (11.24), and keep terms only up to the second, to obtain

+ !(z - ZO)2 I" (zo). Let us assume that t" (zo) f= 0, and define

(11.25)

f(z) = f(zo)

z - zo =

Yje

i B,

and

!f"(zo) =

T2eifh

(11.26)

and substitute in the above expansion to obtain f(z) - f(zo) =

TfT2ei(2Bz+fh),

(11.27)

314

11. COMPLEX ANALYSIS: ADVANCED TOPICS

z,t>~O

y

__ Co

x

Figure 11.12 A segment of thecontour Co in thevicinityof zoo Thelines mentioned in thetextaresmallsegments of thecontour Co centered at zoo

or

+ 02), rrr2 sin( 201 + ~).

Re[f(z) - f(zo)] = rhcos(201 Im[f(z) - f(zo)] =

(11.28)

The constancy of Im[f(z)] implies that sin(201 + 02) = 0, or 201 + 02 = ntt . Thus, for 01 = -~/2 + nIT/2 where n = 0, I, 2, 3, the imaginary part of f is constant. The angle 02 is determined by the second equation in (11.26). Once we determine n, the path of saddle point integration will be specified. To get insight into this specification, consider z - zo = rlei(-1J,/2+nrr/2), and eliminate rl from its real and imaginary parts to obtain

y-yo=[tanC;

method ofsteepest descent

-~)](X-xo).

This is the equation of a line passing through zo = (xo, YO) and making an angle of 01 = (nIT - ~)/2 with the real axis. For n = 0,2 we get one line, and for n = 1,3 we get another that is perpendicular to the first (see Figure 11.12). It is to be emphasized that along both these lines the imaginary part of fez) remains constant. To choose the correct line, we need to look at the real part of the function. Also note that these "lines" are small segments of (or tangents to) the deformed contourat zoo We are looking for directions along which Re(f) goes through a relative maximum at zoo In fact, we are after a path on which the function decreases maximally. This occurs when Re[f(z)] - Re[f(zo)] take the largest negative value. Equation (11.28) determines such a path: It is that path on which cos(201 + 02) = -I, or when n = 1,3. There is ouly one such path in the region of interest, and the procedure is uniquely determined/' Because the descent from the maximum value at

11.5 METHOO OF STEEPEST OESCENT

315

zo is maximum along such a path, this procedure is called the method of steepest descent. Now that we have determined the contour, let us approximate the iutegral. Substituting 2ej + e2 = tt, 3n in Equation (11.27), we get fez) - f(zo)

= -rir: es _t2 = i(z -

ZO)2 f"(zo).

(11.29)

Using this in Equation (11.23) yields

I (a) "" ( ea[i(zoH'] g(z) dz = eai C,o) ( e-at' g(z) dz,

leo

leo

(11.30)

where Co is the deformed contour passing through zoo To proceed, we need to solve for z in terms of t. From Equation (11.29) we have

2

2

t

2

2

(z - zo) = - - - I = --e !"(zo) ri

-ie,

=

=

Therefore, [z - zo I It1/ JT2, or Z - zo (It1/JT2)e ie " by the first equation of (11.26). Let us agree that for t > 0, the point z ou the coutour will move in the direction that makes an angle of 0 :5 el < tt , and that t < 0 correspouds to the opposite direction. This convention removes the remaining ambiguity of the angle el, and gives 1

·01

z=zo+-el

JT2

(lUI)

,

Using the Taylor expansion of g(z) about zn, we can write g(z)dz =

{~~einOlg(n)(zO)} L.- n/2 I n=O 72

L-

'ri.

dt

'V':l.

tn

00

= "

n.

iO, e

(n+l)/2 ,

n=O 72

ei(n+I)01 g(n\ZO) dt I

n.

and substituting this in Equation (11.30) yields I(a) "" eaiC.o) { e- at' o

lee 00

= eaiC'o) "

L.-

e

{~ L.-

n t

Cn+I)/2 , n.

n=O 72

i(n+l)91

Cn+I)/2 , n=O 72 n.

g(n) (ZO)

eiCn+I)O, gcn)(ZO)} dt

JOO e-at't n dt.

(11.32)

-00

5Theangle191 is still ambiguous by x, becausen can be 1 or 3. However, by a suitablesign convention described below,we canremovethis ambiguity.

316

11. COMPLEX ANALYSIS: ADVANCED TOPICS

asymptotic expansion of I(a)

The extension of the integral limits to infinity does not alter the result significantly because a is assumed large and positive. The integral in the sum is zero for odd n. Whenn is even, we make the substitution u = at 2 and show that J~oo e- a t2 t" dt = a-(n+l)/2r[(n+ 1)/2]. Withn = 2k, and using r2 = I f"(zo) 1/2,the sum becomes 2k+l/2 i(2k+I)el

00

[(a) "" ea!(,o) ' "

e

f;;o If"(zo)lk+l/2(2k)!

g(2k) (zo)r(k

+ !)a- k- 1/2. 2

(11.33)

This is called the asymptotic expansion of [(a). In most applications, only the first term of the above series is retained, giving [(ex) "" ea!(,o)

fiIT eiel g(zo) . V-;- JI f" (zo)1

(11.34)

11.5.1. Example. LeI us approximale theintegral [(ex) == r(ex

+ 1) = faOO e-,zadz,

where ex is a positive real number. First, we must rewrite the integral in the fonn of Equation (11.23). We can do this by noting that za ea ln z. Thus,we have

=

l(ex) =

10

00

ealnZ-Zdz =

10

00 elX(lnz-z/a)dz,

andWe identify f(z) = ln z - z/a and g(z) or zu = C1. Furthermore, from 21 f " (zn) and 2(h

Stirling

= 21 ( -

+ 6z =

a1) 2

= I. The saddle point is found from f'(z)

= 0

= 2cx12 e in

x, 3n. as well as the condition 0 ::::: 81 < zr, we conclude that 91 = O.

Substitution in Equation (11.34) yields

approximation r(ex

+ 1) "" ea!(,o)

rzrr_l_

V-;;./1/ex2

= .J2Jrexea(lna-l) = 5 e - aa a+l / 2 , (11.35)

called the Stirling approximation.

III

11.5.2. Example. The Hankel function ofthe first kind is defined as HSI)(ex)

==,!- ( lJr

lc

e(a/2)(,-1/zl..!!:.-, zv+l

where C is the contour shown in Figure 11.13. We want to find the asymptotic expansion of this function, choosing the branch of the function in which -l'C < < 'Jr. We identify f(z) = :!:(z - I/z) and g(z) = z-v-I. Next, the stationary points of f are calculated:

e

df I dz = 2:

I

+ 2z2 = 0

=}

Zo = ±i.

11.5 METHOD OF STEEPEST DESCENT

317

Irnz in

1-----.----o

Rez

Figure 11.13 The contourfor the evaluation of the Hankelfunction of the firstkind. The contour of integration suggests the saddle point ZQ =

+i. The second derivative evalu-

ated at the saddle point gives f"(zo) = -1/z5 = -i = e- i" j2 , or Ih = -,,/2. This, and the convention 0 ::::: 611 < 1r, force us to choose 611 = 3.1l"/ 4. Substituting this in Equation (11.34)and noting that f(i) = i and II"(zo)1 = I, we obtain

H~l)(a) es ~I(a) ~ ~eai fiJre i3rr/ 4i - v- l m

IX

V-;;

= (2ica-vlr/z-n/4),

Van

where we have used i- v - 1 = e- i(v+l)rrj2.

III

Although Equation (11.34) is adequate for most applications, we shall have occasion to demand a better approximation. One may try to keep higher-order terms of Equation (11.33), but that infinite sum is in reality inconsistent. The reason is that in the product g(z) dz; we kept only the first power of t in the expansion of z. To restore consistency, let us expand z(t) as well. Snppose 00

00

z-zo = Lbmtm

=}

dz =

L(m + l)bm+1tmdt, m=O

m=l

so that 00

g(z) dz =

00

n=O 72

n.

L

e i n8r

00

=

tn

L -;;72-/no, g(n) (zo) L (m + I)bm+ltmdt m=O

~(m + l)b m+lg(n)(Zo)t m+ndt.

m,n=O T2

Now introdnce 1 = m gives

n.

+ n and note that the summation over n goes np to I. This

318

11. COMPLEX ANALYSIS: ADVANCED TOPICS

Substituting this in Equation (11.30) and changing the contour integration into the integral from -00 to 00 as before yields 00

I(a) "" ea!(zo) Lazka-k-I/Zr (k

+!),

k~O

2k

ein(h

au = L ---nj2(2k - n + l)bZk_n+lg(n)(zo). n=O r2 n!

(11.36)

The only thing left to do is to evaluate bm • We shall not give a general formula for these coefficients. Instead, we shall calculate the first three of them. This shonld reveal to the reader the general method of approximating them to any order. We have already calculated bt in Equation (11.31). To calculate bz, keep the nexthighest term in the expansion of both z and t Z• Thus write tZ =

-~ f"(zo)(z -

zo)z -

2

~ f"'(zo)(z 6

ZO)3.

Now substitute the first equation in the second and equate the coefficients of equal powers of t on both sides. The second power of t gives nothing new: It merely reaffirms the value of bl. The coefficient of the third power of t is -bl bzt" (zo) ibif"'(zo). Setting this equal to zero gives

bz =

bf fill (zo) _ fill (zo) e4i9[ 6f"(zo) - 31f"(zo)12 ,

(11.37)

where we substituted for bl from Equation (11.31) and used 201 + Oz = tt . To calculate b3, keep one more term in the expansion of both z and t Z to obtain

and 2 1" 2 1 t = - - f (zo)(z - zo) - - f

2

m

6

3 1 (zo)(z - zo) - - f

24

(iv)

(zo)(z - zo)

4

.

Once again substitute the first equation in the second and equate the coefficients of equal powers of t on both sides. The second and third powers of t give nothing new. Setting the coefficient of the fourth power of t equal to zero yields

b _ b3 { 5[f"'(zo)]z f(iV)} 3 I 72[f"(zo)]Z - 24f"(zo)

../ie3i 9[

{5[flll(zo)f

f(iV)}

= 121f"(zo)13/ Z 3[f"(zo)]Z - f"(zo)

.

(11.38)

319

11.6 PROBLEMS

Imz

2i1[;/'

+ __---,

\..LJ •

f\ -00

o

Figure 11.14

Contourused for Problem 11.4.

+00

Rez

11.6 Problems 11.1. Derive Equation (11.2) from its logarithmic derivative. 11.2. Show that the point at infinity is not a branch point for f(z) = (z2 - 1)1/2. 11.3. Find the following integrals, for which 0

t= a E R. 00

(e)

1

(lnx)2 22dx.

o x +a

11.4. Use the contour in Figure 11.14 to evaluate the following integrals. (a)

roo s~ax dx

Jo

(b)

roo x ~osax dx

x Jo smhx Show that 1; f(sine) de = 21;/2 f(sine)de for an arbitrary function f sinh

11.5. defiued iu the interval [-I, +1].

11.6. Find the principal value of the integral 1.':"00 x sin x dx / (x 2- x5) and evaluate 1=

1

00

-00

x sinx d x (x - xo ± i€)(x + Xo ± i€)

for the four possible choices of signs.

11.7. Use analytic continuation, the analyticity of the exponential, hyperbolic, and trigonometric functions, and the analogous identities for real z to prove the following identities. (a) eZ = cosh z + sinh z. (e) sin2z = 2sinzcosz.

(b) cosh2 z - sinh2 Z = 1.

320

11. COMPLEX ANALYSIS: ADVANCED TOPICS

11.8. Show that the function I/z 2 represents the analytic continnation into the domain iC - {OJ (all the complex plane minns the origin) of the function defined by L:~o(n + I)(z + I)" where [z + II < 1. 11.9. Find the analytic continuation into iC- Ii, -i} (all the complex plane except oo i and -i) of f(z) = fo e- zt sin t dt where Refz) > O. 11.10. Expand f(z) = L:~o zn (defined in its circle of convergence) in a Taylor series about z = a. For what values of a does this expansion permit the function f (z) to be continued analytically? 11.11. The two power series 00

Ji(z) =

zn

Ln=l n

. !z(z) = in

and

~

+ L..(-I)

n (z _2)n -'--'--------'-

n

n=l

have no common domain of convergence. Show that they are nevertheless analytic continuations of one another. 11.12. Prove that the functions defined by the two series

and

I (I - a)z (I - a)2 z2 -I---z - (I - Z)2 + -'-(:-:-I-_--"Z"")3c-

are analytic continuations of one another. 11.13. Show that the function Ji (z) = I/(z2 + I), where z i' ±i, is the analytic continuation into iC - Ii, -i} of the function !z(z) = L:~0(_l)nz2n, where lel < 1. 11.14. Find the analytic continuation into iC - {OJ of the function

f(z) =

10

00

te- ztdt

where Retz) > O.

11.15. Show that the integral in Equation (11.9) converges. Hint: First show that [F'(z + I) I ::s fooo t X e- t dt where x = Retz), Now show that

for some integer n > 0 and conclude that I'(z) is finite. 11.16. Show that dr(z + 1)ldz exists and is finite by establishing the following: (a) [Inr] < t+ lit for t > O. Hint: Fort::: I, showthatt -Int is a monotonically increasing function. For t < I, make the substitution t = lis. (b) Use the result from part (a) in the integral for dr(z + 1)ldz to show that Idr(z + I)ldzl is finite. Hint: Differentiate inside the integral.

11.6 PROBLEMS

321

11.17. Derive Equation (11.11) from Equation (11.9). 11.18. Show that I' (~) =

.ft, and that k

(Zk - I)l!

== (Zk -j)(2k -

2 f .ft

3) .. ·5·3· I =

(2k+l) -Z- .

11.19. Show that F'(z) = Io1[1n(llt)y-ldt with Re(z) > O. 11.20. Derive the identity

10

00

eX" dx = f[(a

+ 1)la].

11.21. Consider the function f (z) = (I + z)". (a) Show that d" fldznlz~o = I'(o + 1)1 I'(o - n relation

a) where ( n

==

+ I), and use it to derive the

a! nl(a - n)!

==

f(a + I) n!r(a - n + I) .

(b) Show that for general complex numbers a and b we can formally write

(a + b)a =

t

(:)anb

a-n.

n=O

(c) Show that if a is a positive integer m, the series in part (b) truncates at n = m. 11.22. Prove that the residue of I'(z) at z = -k is rk = (-I)klkL Hint: Use Equation (11.12) 11.23. Derive the following relation for z = x

+ iy:

11.24. Using the definition of B(a, b), Equation (11.16), show that B(a, b) = B(b, a). 11.25. Integrate Equation (11.21) by parts and derive Equation (11.11). 11.26. For positive integers n, show that f(~ - n)r(~ 11.27. Show that

(a) B(a, b) = B(a

+ I, b) + B(a, b + I).

(b) B(a, b + I) =

C~

b) B(a, b).

(c) B(a, b)B(a + b, c) = B(b, c)B(a, b + c).

+ n) =

(_1)n"..

322

11. COMPLEX ANALYSIS: ADVANCED TOPICS

Imz

o

Rez

-it: 1 - - - - -.....- - - The contour for the evaluation of the Hankel function of the second kind.

Figure 11.15

11.28. Verify that

12 1(1 + t)a(l -

t)bdt = za+b+l B(a

+ 1, b + 1).

l,

the 11.29. Show that the volume of the solid formed by the surface z = x a xy-, yz-, and xz-planes, and the plane parallel to the z-axis and going through the points (0, YO) and (xo, 0) is xa+1 b+l

o Yo B(a + 1, b + 1). a+b+Z 11.30. Derive this relation: where - 1 < a < b. Hint: Let t = tanh 2 x in Equation (11.16). 11.31. The Hankel function of the second kind is defined as

where C is the contour shown in Figure 11.15. Find the asymptotic expansion of this function. 11.32. Find the asymptotic dependence ofthe modified Bessel function of the first kind, defined as . [v(ex) es

~

1. e(a/2)«+I/<) dz , zv+l Y

2:rrz c

where C starts at -00, approaches the origin and circles it, and goes back to -00. Thus the negative real axis is excluded from the domain of analyticity.

11.6 PROBLEMS

323

11.33. Find the asymptotic dependeoce of the modified Bessel function of the second kind:

K (a) IJ

==! [

2 Jc

e-(aI2)(z+1/z)~

zv+l'

where C starts at 00, approaches the origin and circles it, and goes back to Thus the positive real axis is excluded from the domain of analyticity.

00.

Additional Reading I. Denoery, P. and Krzywicki, A. Mathematics for Physicists, Harper and Row, 1967. 2. Lang, S. Complex Analysis, 2nd ed., Springer-Verlag, 1985. Contains a very lucid discussion of analytic continuation.

Part IV

Differential Equations

_

12

_

Separation of Variables in Spherical Coordinates

The laws of physics are almost exclusively writteu in the form of differential equations (DEs). In (point) particle mechanics there is only one independent variable, leading to ordinary differential equations (ODEs). In other areas of physics in which extended objects such as fields are studied, variations with respect to position are also important. Partial derivatives with respect to coordinate variables show up in the differential equations, which are therefore called partial differentiai eqnations (PDEs). We list 'the most common PDEs of mathematical physics in the following.

12.1 PDEs of Mathematical Physics Poisson's equation

In electrostatics, where time-independent scalar fields such as potentials and vector fields such as electrostatic fields are studied, the law is described by Poisson's equation,

V2<1>(r) = -41fp(r). Laplace's equation

(12.1)

In vacuum, where p(r) = 0, Equation (12.1) reduces to Laplace's equation, (12.2)

heat equation

Many electrostatic problems involve couductors held at constant potentials and situated in vacuum. In the space between such conducting surfaces, the electrostatic potential obeys Equation (12.2). The most simplified version of the heat equation is (12.3)

328

12. SEPARATION OF VARIABLES IN SPHERICAL COOROINATES

wave equation

where T is the temperature aud a is a constaut characterizing the medium io which heat is flowing. One of the most frequently recurriog PDEs encountered in mathematical physics is the wave eqnation, (12.4)

This equation (or its simplification to lower dimensions) is applied to the vibration of striogs aud drums; the propagation of sound io gases, solids, aud liquids; the propagation of disturbauces in plasmas; aud the propagation of electromagnetic waves.

The Schriidinger equation, describiog nonrelativistic quautum phenomena, Schrodinger equation

is ~ - - V2 2m

Klein-Gordon

w+ V(r)W =

aw -Ui-, at

(12.5)

where m is the mass of a subatomic particle, Ii is Plauck's constaut (divided by 2".), V is the potential energy of the particle, aud IW(r, t)I 2 is the probability density of findiog the particle at r at time t. A relativistic generalization of the Schrodinger equation for a free particle of mass ,m is the Klein-Gordon eqnation, which, in terms of the natural units (Ii = 1 = c), reduces to

equation 2 2 a2q, Vq,-mq,=-2'

at

time isseparated from space

(12.6)

Equations (12.3-12.6) have partial derivatives with respect to time. As a first step toward solviog these PDEs aud as au iotroduction to similar techniques used io the solution of PDEs not iovolviog time, 1 let us separate the time variable. We will denote the functions io all four equations by the generic symbol W(r, t). The basic idea is to separate the r aud t dependence ioto factors: W(r, t) es R(r)T(t). This factorization permits us to separate the two operations of space differentiation aud time differentiation. Let L staud for all spatial derivative operators aud write all the relevaut equations either as LW = aw jat or as L\j1 = a2w jat 2 • With this

notation andthe above separation, we have L(RT) = T(LR) = {RdTjdt, Rd2Tjdt 2

1See [Hass99] for a thorough discussionof separation in Cartesian andcylindrical coordinates. Chapter 19 of thisbook also contains examples of solutions to some second-order linear DEs resulting from suchseparation.

12.1 PDES DF MATHEMATICAL PHYSICS

329

Dividing both sides by R T, we obtain

(12.7)

Now comes the crucial step in the process of separation of variables. The LHS of Equation (12.7) is a function of position alone, and the RHS is a function of time alone. Since r and t are iudependent variables, the only way that (12.7) can hold is for both sides to be constant, say a:

1 -LR = a R

1 dT --=a T dt

=}

LR = «R

dT -=aT dt

=}

We have reduced the original time-dependent POE to an ODE,

dT =aT dt

(12.8)

or

and a POE involving only the position variables, (L - a)R = O. The most general formofL-a arising from Equations (12.3-12.6) is L-a sa V2 + j(r). Therefore, Equations (12.3-12.6) are equivalent to (12.8), and V 2R

+ j(r)R =

O.

(12.9)

To include Poisson's equation, we replace the zero on the RHS by g(r) == -41!'p(r), obtaining V 2R + j(r)R = g(r). With the exception of Poisson's equation (an inhomogeneous POE), in all the foregoing equations the term on the RHS is zero.' We will restrict ourselves to this so-called homogeneous case and rewrite (12.9) as

V2W(r) + j(r)W(r) = O.

(12.10)

Depending on the geometry of the problem, Equation (12.10) is further separated into ODEs each involving a single coordinate of a suitable coordinate system. We shall see examples of all major coordinate systems (Cartesian, cylindrical, and 210 most cases, a is chosen to be real. In the case of the Schrodinger equation, it is more convenientto choose a to be purely imaginary so thatthe i in the definition of Lcanbe compensated. In all cases,theprecisenature of a is determined byboundary conditions. 3Techniques for solvinginhomogeneous PDEsarediscussed in Chapters 21 and22.

330

12. SEPARATION OF VARIABLES IN SPHERICAL COORDINATES

spherical) in Chapter 19. For the rest of this chapter, we shall concentrate on some general aspects of the spherical coordinates.

Jean Le Rond d'A1embert (1717-1783) was the illegitimate son of a famous salon hostess of eighteenth-century Paris and a cavalry officer. Abandoned by his mother, d' Alembert was raised by a foster family and later educated by the arrangement of his father at a nearby church-sponsored school, in which he received instruction in the classics and above-average instruction in mathematics. After studying law and medicine, he finally chose to pursue a career in mathematics. In the 17408 he joined the ranks of the philosophes, a growing group of deistic and materialistic thinkers and writers who actively questioned the social and intellectual standards of the day. He traveled little (he left France only once, to visit the court of Frederick the Great), preferring instead the company of his friends in the salons, among whom he was well known for his wit and laughter. D' Alembert turned his mathematical and philosophical talents to many of the outstanding scientific problems of the day, with mixed success. Perhaps his most famous scientific work, entitled 'Iraite de dynamique, shows his appreciation that a revolution was taking place in the science of mechanics-the fonnaIization ofthe principles stated by Newton into a rigorous mathematical framework. The philosophy to which d' Alembert subscribed, however, refused to acknowledge the primacy of a concept as unclear and arbitrary as "force," introducing a certain awkwardness to his treatment and perhaps causing him to overlook the important principle

of conservation of energy. Later, d'Alembertproduced a treatise on fluid mechanics (the priority of which is still debated by historians), a paper dealing with vibrating sltings (iu which the wave equation makes its first appearance in physics), and a skillful treatment of celestial mechanics. D' Alembert is also credited with use of the first partial differential equation as well as the first solution to such an equation using separation of variables. (One should be careful interpreting ''first'': many of d' Alembert's predecessors and contemporaries gave similar. though less satisfactory. treatments of these milestones.) Perhaps his most well-known contribution to mathematics (at least among students) is the ratio test for the convergence of infinite series. Much of the work for which d' Alembert is remembered occurred outside mathematical physics. He was chosen as the science editor of the Encyclopedie, and his lengthy Discours Preliminaire in that volume is considered one of the defining documents of the Enlightenment. Other works included writings on law, religion, and music. Since d' Alembert's final years were not especially happy ones, perhaps this account of his life should eud with a glimpse at the bwnanity his philosopby often gave his work. Like many of his contemporaries, he considered the problem of calculating the relative risk associated with the new practice of smallpox inoculation. which in rare cases caused the disease it was designed to prevent. Although not very successful in the mathematical sense. he was careful to point out that the probability of accidental infection, however slight or elegantly derived, would be small consolation to a father whose child died from the

12.2 SEPARATION OFTHEANGULAR PARTOFTHE LAPLACIAN

331

inoculation. It is greatly to his credit that d' Alembert did not believe such considerations irrelevant to the problem.

12.2 Separation of the Angular Part of the Laplacian With Cartesian and cylindrical variables, the boundary conditions are important in determining the nature of the solutions of the ODE obtained from the POE. In almost all applications, however, the angular part of the spherical variables can be separated and studied very generally. This is because the angular part of the Laplacian in the spherical coordinate system is closely related to the operation of rotation and the angular momentum, which are independent of any particular situation. The separation of the angular part in spherical coordinates can be done in a fashion exactly analogous to the separation of time by writing Ij1 as a product of three functions, each depending on ouly one of the variables. However, we will follow an approach that is used in quantum mechanical treatments of angular momentum. This approach, which is based on the operator algebra of Chapter 2 and is extremely powerful and elegant, gives solutions for the angular part in closed form. Define the vector operator p as p = -iV so that its jth Cartesian component is Pj -i8/8xj, for j 1,2,3. In quantum mechanics p (multiplied by Ti) is the momentum operator. It is easy tu verify that" [Xj, Pk] = i8jk and [Xj, xk] = 0= [Pj' Pk]· We can also define the angular momentum operator as C = r x p. This is expressed in components as Li = (r X P)i = EijkXjPk for i = 1,2,3, where Einstein's summation convention (summing over repeated indices) is utilized.f Using the commutation relations above, we obtain

=

angular momentum operator commutation relations between components 01 angular momentum operator

=

[Lj, Lk] = iEjk/L/.

Ccan be written solely in terms of the angles e and p in the definition of C, so if we square C, we will get two factors of p, and a Laplacian may emerge in the expression for C. L. We will see shortly that

tp, Moreover, there is one factor of

In this marmer, we may be able to write '11 2 in terms of L2 , which depends only on

4These operators act on the space of functions possessing enough "nice" properties as to render the space suitable. The operator x j ~ply multiplies functions, while p' differentiates them.

It IS assumed that the reader is familiar with vector algebra using indices and such objects as 8ij and Eijk' For an introductory treatment, sufficient for our present discussion, see [Hass 99]. A more advanced treatment of these objects (tensors) can be found in Part VII of this book.

332

12. SEPARATION OF VARIABLES IN SPHERICAL COORDINATES

angles. Let us try this: 3 " L2 = " L·" L = .!...J LiL; =

= EijkEimn X jPkXmPn

EijkX jPkEimllXmPn

;=1

= (8jm8kn - 8jn8km)XjPkXmPn = XjPkXjPk - XjPkXkPj'

We need to write this expression in such a way that factors with the same index are next to each other, to give a dot product. We must also try, when possible, to keep the pfactors to the right so that they can operate on functions without intervention from the x factors. We do this using the commutation relations between the x's and the p's: 2 L =

Xj(XjPk - i8kj)Pk -

(PkXj

+ i 8kj)XkPj

= XjXjPkPk - iXjPj - PkXkXjPj - iXjPj = XjXjPkPk - 2ixjPj - (XkPk - i8kk)XjPj'

Recalling that 8kk = Ei~l 8kk = 3 and XjXj = E]~l XjXj = r· r = r 2 etc., we can write L 2 = r 2 p.p+ rr· p- (r. p)(i'. p), which, if we make the substitution p = -iV, yields V 2 = _r- 2 L2

+ r- 2(r. V)(r· V) + r-2r · V.

Letting both sides act on the function "'(r, e,
2

I 2 r

1 r

V '" = - -2L '" + -(r· V)(r· V)'" 2

Laplacian separated into angular and radial parts

But we note that r· V spherical coordinates: V2", =

1

+ -r 2r · V"'.

(12.11)

= reT' V = ra/ar. We thus get the final form of \72", in

_~L2", + ~~ (r a"') + ~ a",. 2

r

r ar

ar

(12.12)

r ar

It is important to note that Equation (12.11) is a general relation that holds in all.coordinate systems. Although all the manipulations leading to it were done in Cartesian coordinates, since it is written in vectornotation, there is no indication in the final form that it was derived using specific coordinates. Equation (12.12) is the spherical version of (12.11) and is the version we shall use. We will first make the simplifying assumption that in Equation (12.10), the master equation, fer) is a function of r only. Equation (12.10) then becomes

1

2

--L '" 2

r

1 a (a"') I a", + -- + f(r)'" r ar rar- + r ar

= O.

e

Assuming, for the time being, that L2 depends only on and
12.2 SEPARATION OFTHEANGULAR PARTOFTHE LAPLACIAN

333

equation as l a [r-(RY) a ]+ I--(RY) a + -+ f(r)RY = r ar ar r ar

- -I2L2(R Y)

r

O.

Dividing by RY and multiplying by r 2 yields

I Y

2

r d (dR) Rdr dr

r dR R dr

2

--L (Y)+-- r - + - - + r f(r) =0, " - . . - '



I

+a

-no

or (12.13) and 2R d 2dR [ Ci] '-+--+ f(r)--2 R=O. dr 2 r dr r

(12.14)

We will concentrate on the angular part, Equation (12.13), leaving the radial part to the general discussion of ODEs. The rest of this subsection will focus on showing that Ll sa Lx> L2 sa Ly , and L3 sa Lz are independent of r. Since Li is an operator, we can study its action on an arbitrary function f. Thus, L;j = -i€ijkX/'hf == -i€ijkXjaf/aXk. We can express the Cartesian Xj in terms of r, II, and tp, and use the chain rule to express af/axk in terms of spherical coordinates. This will give us Li f expressed in terms of r, II, and rp. It will then emerge that r is absent in the final expression. Let us start with x = r sin IIcos tp, Y = r sin IIsin tp, z = r cos II, and their inverses, r = (x 2 + y 2 + Z2)1/2, cosll = zfr, tanrp = Ylx, and express the Cartesian derivatives in terms of spherical coordinates using the chain rule. The first such derivative is

af ax

af ar ar ax

af all all ax

af arp arp ax

-=--+--+--.

(12.15)

The derivative of one coordinate system with respect to the other can be easily calculated. For example, arlax = x]r = sinllcosrp, and differentiating both sides of the equation cos II = zl r, we obtain

. all zarlax zx -smll- = - - - = -= 2 3

ax

r

cos IIsin 1/ cos rp

r

r

=}

all cos IIcos rp = ---'ax r

Finally, differentiating both sides oftan rp = YI x with respectto x yields arp I ax = - sinrp/(r sin II). Using these expressions in Equation (12.15), we get

af. af = smllcosrpax ar

-

+ cosllcosrpaf - -sinrp - -af. r

all

r sin II arp

334

12. SEPARATION OF VARIABLES IN SPHERICAL COOROINATES

In exactly the same way, we obtain

af . . af cos e sin I" af cos I" af -=smesmqJ,,+ oe+-'-e'" ay or r u r sm vI" af af sine af - = cos()- - - - - . az ar r ae We can now calculate Lx by letting it act on an arbitrary function and expressing all Cartesian coordinates and derivatives in terms of spherical coordinates. The result is

. af + IZ. af = I. (smqJ. a + cote cos 1"a) Lxf = -lyaz ay ae aqJ f, Cartesian components of angular momentum operator expressed inspherical coordinates

or Lx = i (Sin qJ~

ae + cote cos qJ~) aqJ .

Analogous arguments yield Ly = i

angular momentum squared as differential operator in e and q>

(12.16)

(-cosqJ~ ae + cote SinqJ~), aqJ

(12.17)

It is left as a problem for the reader to show that by adding the squares of the components of the angular momentum operator, one obtains I L = - sine 2

a (. a ) ae sme ae -

2

I a sin 2 e aqJ2'

(12.18)

Substitution in Equation (12.12) yields the familiar expression for the Laplacian in spherical coordinates.

12.3 Construction of Eigenvalues of l2 Now that we have L2 in terms of e and 1", we could substitute in Equation (12.13), separate the and I" dependence, and solve the corresponding ODEs. However, there is a much more elegant way of solving this problem algebraically, because Equation (12.13) is simply an eigenvalue equation for L2. In this section, we will find the eigenvalues of L2 • The next section will evaluate the eigenvectors of L2 Let us consider L2 as an abstract operator and write (12.13) as

e

where IY) is an abstract vector whose (e, qJ)thcomponent can be calculated later. Since L2 is a differential operator, it does not have a (finite-dimensional) matrix

12.3 CONSTRUCTION OF EIGENVALUES OF L 2

335

representation. Thus, the determinantal procedure for calculating eigenvalues and eigenfunctions will not work here, and we have to find another way. The equation above specifies an eigenvalue, a, and an eigenvector, IY). There may be more than one IY) corresponding to the same a. To distinguish among these so-called degenerate eigenvectors, we choose a second operator, say L3 E {L,}that commntes with L2. This allows us to select a basis in which both L2 and L3 are

diagonal, or, equivalently, a basis whose vectors are simultaneous eigenvectors of both L2 and L3. This is possible by Theorem 4.4.15 and the fact that both L2 and L3 are hermitian operators in the space of square-integrable functions. (The proof is left as a problem.) In general, we would want to continue adding operators until we obtained a maximum set of commuting operators which could label the eigenvectors. In this case, L2 and L3 exhaust the set," Using the more common subscripts x, y, and z instead of 1, 2, 3 and attaching labels to the eigenvectors, we have Lz IY.,p) = fJ 1Y.,p),

(12.19)

The hermiticity of L2 and Lz implies the reality of a and fJ. Next we need to determine the possible values for a and fJ. Define two new operators L+ sa Lx + i Ly and L == Lx - i Ly . It is then easily verified that (12.20) The first equation implies that L± are invariant operators when acting in the subspace corresponding to the eigenvalue a; that is, L± 1Y.,p) are eigenvectors of L2 with the same eigenvalue a: L2(L± 1Y.,p) = L±(L21Y.,p) = aL± 1Y.,p).

The second equation in (12.20) yields Lz(L+ IY.,p)

= (LzL+) 1Y.,p) = (L+Lz + L+) IY.,p) = L+Lz 1Y.,p) + L+ lYa,p) = fJL+ 1Ya,p) + L+ 1Y.,p) = (fJ + I)L+ IY.,p) .

angular momentum raising and lowering operalors

This indicates that L+ 1Y.,p) has one more unit of the Lz eigenvalue than 1Y.,p) does. In other words, L+ raises the eigenvalue of Lz by one unit. That is why L+ is called a raising operator. Similarly, L is called a lowering operator because Lz(L 1Y.,p) = (fJ - I)L 1Ya,p).

Wecan summarize the abovediscussion as

6We couldjust as well have chosen L2 andanyother component as ourmaximalset. However, L2 and L3 is the universally accepted choice.

336

12. SEPARATION OF VARIABLES IN SPHERICAL COOROINATES

where C± are constants to be determined by a suitable normalization. There are restrictions on (and relations between) 01 and p. First note that as L2 is a sum of squares of hermitian operators, it must be a positive operator; that is, (al L2 1a) :::: 0 for all ]c). In particular, 0::; (Ya,~1 L2IYa,~) =

Therefore,

01 :::: O.

2 L = L+L

01 (Ya,~1

Ya,~) = 0IIIYa,~1I2.

Next, one can readily show that

+ L; -

Lz = LL+

+ L; + Lz.

(12.21)

Sandwiching both sides of the first equality between IYa,~) and (Ya,~ I yields

(Ya,~1 L2IYa,~) = (Ya,~1 L+L IYa,~) + (Ya,~1 L; lYa,~) - (Ya,~1 t, lYa,~), with an analogous expression involving LL+. Using the fact that L+ = (L)t, we get

0IIIYa,~1I2 = (Ya,~1 L+L IYa,~) + p2I1Ya,~1I2 - PIIYa,~1I2 = (Ya,~1 LL+ IYa,~) + p2I1Ya,~1I2 + PIIYa,~1I2 = liLT IYa.~) 11 2 + p2I1Ya,~ 1= PllYa,~1I2

f

(12.22)

Because of the positivity ofnonns, this yields 01 :::: p2 - P and 01 :::: p2 + p. Adding these two inequalities gives 201 :::: 2p2 => -v'{i::; p ::; v'{i. It follows that the values of p are bounded. That is, there exist a maximum p, denoted by P+, and a minimum p, denoted by p_, beyond which there are no more values of p. This can happen only if

because if L± lYa,~±) are not zero, then they must have values of p corresponding to P± ± 1, which are not allowed. Using P+ for p in Equation (12.22) yields (01- p~

- P+)llYa,~+1I2 =

O.

By definition lYa,~+) oft 0 (otherwise P+ - 1 wonld be the maximum). Thus, we obtain 01 = p~ + P+ An analogous procedure using p_ for p yields 01 = P: - p_. We solve these two equations for P+ and p_:

P+

=

Since P+

P+

!(-I ±"'1 +4a),

p_ =!(I ±

"'I

:::: p_ and "'1 + 401 :::: I, we must choose

= !(-I + "'1 + 401) = -p_.

+4a).

12.3 CONSTRUCTION OF EIGENVALUES OF L 2

337

Starting with 1Ya.~+), we can apply L_ to it repeatedly. In each step we decrease the value of fJ by one unit. There must be a limit to the number of vectors obtained in this way, because fJ has a minimum. Therefore, there must exist a nonnegative integer k such that

(L)k+1IYa,~+)

= L(L~ lYa,~+)) = O.

Thus, L~ lYa,~+) must be proportioual to lYa,~_). In particular, since L~ IYa,~+) has a fJ value equal to fJ+ - k, we have fJ- = fJ+ - k. Now, using fJ- = -fJ+ (derived above) yields the important result

k . fJ+=Z=J

forkEN,

or a = j (j + 1), since a = fJt + fJ+. This result is important enough to be stated as a theorem. eigenvalues of L' and Lz given

12.3.1. Theorem. The eigenvectors of~, denoted by 1Yjm), satisfy the eigenvalue relations 21 L Yjm) = j(j + 1) IYjm) ,

L z IYjm) = m IYjm) ,

where j is a positive integer or half-integer, and m can take a value in the set I-j, - j + 1, ... ,j -1, j] of2j + 1 numbers. Let us briefly consider the normalization of the eigenvectors. We already know that the IYjm), being eigenvectors of the hermitian operators L2 and Lz, are orthogonal. We also demand that they be of unit norm; that is,

(12.23) This will determine the constants C±, introduced earlier. Let ns consider C+ first, which is defined by L+ IYjm) = C+ IYj,m+l). The hermitian conjngate of this equation is (Yjml L = C':' (Yj,m+ll. We contract these two eqnations to get (Yjml LL+ IYj m) = IC+1 2 (Yj,m+ll Yj,m+l). Then we use the second relation in Equation (12.21), Theorem 12.3.1, and (12.23) to obtain j (j

+ 1) -

m (m + 1)

= 1C+ 12 =>

1C+ 1 = -/"j-;-(j:---+:---I:7)---m:---(:---m-+:---I'"'").

Adopting the convention that the argument (phase) of the complex number C+ is zero (and therefore that C+ is real), we get

C+ = .jj(j + 1) -m(m + 1)

+ 1) L+ IYjm) = ,jj(j + 1) L IYjm) =,j j(j + 1) -

Similarly, C- = -/ j (j

m(m

1). Thus, we get

m(m + 1) IYj,m+l) , m(m - 1) IYj,m-l).

(12.24)

338

12. SEPARATION OF VARIABLES IN SPHERICAL COOROINATES

12.3.2. Example. Let us find an expressionfor IYlm) by repeatedly applyingL- to IYu). The actionfor L_ is completely described by Equation (12.24). Forthe first power of L_, we obtain L-lYu)

= ,//(/ + I) -/(/-1) IYI,I-I} = J2l1Y1,l-1).

Weapply L_ oncemore: (L-)2IYU) = J2lL-1Y1,1-l} = J2l,//(l

+ I) -

(/- 1)(/- 2) IYI,I-2}

= J2l,/2(21 - I) 1Y1,1-2} = ,/2(21)(2/ - I) 1Y1,1-2)'

ApplyingL- a third time yields (L-)3 IYU) = ,/2(2/)(2/- 1)L-IYI,l-2} = ,/2(21)(2/- 1),/6(/- I) 1Y1,1-3) = ,/3!(21)(21 - 1)(21 - 2) IYI,H}.

Thepattern suggeststhefollowing formula for a generalpowerk:

L~ lYu) = ,/k!(2/)(21-I) ... (2/- k + I) 1Y1,I-k},

or L~ lYu} IYl,m }

= ,/k!(21)!j(21 (I

+ m)!

k)! 1YI,I-k).lfwe set 1- k

= m and solvefor IYI,m)' we get

l-mIY)

= (/ _ m)!(21)! L_

iu-

III

The discussion in this section is the standard treatment ofangular momentum in quantum mechanics. In the context of quantum mechanics, Theorem 12.3.1 states the far-reaching physical result that particles can have integer or half-integer spin. Suchaconclusion is tiedtotherotation group in three dimensions, which,in tum, is an exaruple ofa Lie group, or a contiuuous group of transformations. We shall come back to a study ofgroups later. It is worth notiug that it was the study of differential equations that led the Norwegian mathematician Sophus Lie to the investigation of their symmetries and the development of the beautiful branch of mathematics and theoretical physics that bears his narue. Thus, the existence of a conuection between group theory (rotation, angular momentum) and the differential equation we are trying to solve should not come as a surprise.

12.4 Eigenvectors of L2 : Spherical Harmonics The treatment in the preceding section took place in an abstract vector space. Let us go back tu the fuuction space and represent the operators and vectors in terms of 0 and
12.4 EIGENVECTORS OF L 2: SPHERICAL HARMONICS

339

We write Yjm(O,
In classical physics the value of functions must be the same at


12.4.1. Box. From now on, we shall assume that m is an integer and denote the eigenvectors of L2 by Ylm(0,
Our task is to find an analytic expression for Ylm(O,
e±i~ (±~ + i cotO~) . ao

(12.25)

a
Since 1is the highest value ofm, when L+ acts on Yll(O,
(:0 + I cot 0aa
=}

(:0 -I cot 0) Pll(O) = O.

The solution to this differential equation is readily found to be Pll (IJ) = CI (sin 0)1.

The constant is subscripted because each Pll may lead to a different constant of integration. We can now write Yll(O,
. ( a

a ) [Pll(O)e'~] '1

LYli = e-'~ - - + I cotOao a
:0 + I

=

e-i~ [-

=

(_I)ei(l-l)~ (:0

cotO(il)]

[Pll(O)eil~]

+lcotO) Pll(O).

340

12. SEPARATION OF VARIABLES IN SPHERICALCOORDINATES

It can be shown that for a positive integer, d ) I d .n ( dO +ncotO f(O) = sinnOdO[sm Of(O)].

(12.26)

Using this result yields LYIl =

=

v'2i YI.I-l

= v'2iei(l-ll~ PI,1-1 (0)

(_I)ei(I-ll~_I_~[sini O(CI sini 0)] siniO dO d

ei(I-l)~

= (-I)CI---(sin21 0). sinlO dO

(12.27)

We apply L to (12.27), and use Equation (12.26) with n = I - I to obtain

With a little more effort one can detect a pattern and obtain k

dk

e i (l -k) ~

(

L_ Yn = CI (1 _ u 2)(l-kl(2 du k [ I - u

2)1

].

If we let k = I - m and make use of the result obtained in Example 12.3.2, we obtain YIm(O, rp) =

eim~ dI- m (I + m)! [(I _ u2)I ] (1- m)!(21)! I (I _ u 2)m(2 du' m .

--,-'---'-c-=,...,C

To specify Ylm(O, rp) completely, we need to evaluate CI. Since CI does not depend on m, we set m = 0 in the above expression, obtaining l

YlO(U, rp) =

I d 2 I IMmCIdl [(1 - u ) ].

v (21)!

u

The RHS looks very much like the Legendre polynomials of Chapter 7. In fact, YlO(U, rp) =

~(-1)1211!PI(U) '" AIPI(U).

v (21)!

(12.28)

12.4 EIGENVECTORS OF L 2: SPHERICAL HARMONICS

341

Therefore, the normalization of YIO and the Legendre polynomials Pi determines Ct. We now use Equation (6.9) to obtain the integral form of the orthonormality relation for Ylm :

(12.29) which in terms of u = cos e becomes

it 1 2"

o

dip

-t

(12.30)

Yt;m'(U'
Problem 12.15 shows thatnsing (12.29) one gets At = .)(21 + l)j(4n). Therefore, Equation (12.28) yields not only the value of Ci, but also the useful relation YIO(U, '1') = spherical harmonics

r:

(12.31)

--Pt(u). 4n

Substituting the value of Ct thus obtained, we finally get

where u = cos e. These functions, the eigenfunctions of L2 and Lz, are called spherical harmonics. They occur frequently in those physical applications for which the Laplacian is expressed in terms of spherical coordinates. One can immediately read off the part of the spherical harmonics:

e

Ptm(u) = (-1)

associated Legendre functions

t('ii+ll V~ 2tl!

However, this is not the version used in the literature. For historical reasons the associated Legendre functions Pt(u) are used. These are defined by m

Pt (u) = (-1)

m

(I+m)! ~ Ptm(u ) (1- m)!V

2i+1

I + )1 (1 = (_I)t+m ( m. (1- m)! Thus,

2)-m/2 d t- m

u

»n

du' m

[(1-

u2iJ.

i:: :;:t pt"'
(12.33)

2

Ytm(e, '1') = (_1)'" [21 1 4:

(12.34)

342

12. SEPARATION OF VARIABLES IN SPHERICAL COOROINATES

We generated the spherical harmonics starting with Yl!(0,
IYI,-",)

=

(I + m)! LI - ", IY ) (I _ m)!(21)! + 1,-1·

(12.35)

(This can be obtained following the steps of Example 12.3.2,) Next, we use L IYI,-I) = 0 in differential form to obtain

(:0

-I cot 0) PI,-I(O) = 0,

which has the same form as the differential equation for Pl!. Thus, the solution is PI,_I(O) = Cf(sinO)I, and YI,_I(O,
= PI,_I(O)e-il~ = C;(sinO)le-il~.

Applying L+ repeatedly yields (-lle-i(l-k)~ d k

2 I L+YI,_I(U,
k

I

where U = cos O. Substituting k = 1 - m and using Equation (12.35) gives YI,-",(U,
The constant Cf can be determined as before. In fact, for m = 0 we get exactly the same result as before, so we expect Cf to be identical to CI. Thus,

Comparison with Equation (12.32) yields YI,_",(O,
(12.36)

andusing thedefinition Pic.s.Ie',
+

(12.37)

12.4 EIGENVECTORS OF L 2: SPHERICAL HARMONICS

343

The first few spherical harmonics with positive m are given below. Those with negative m can be obtained using Equation (12.36). I

For 1= 0,

Yoo =

Fori = I,

YlO

For 1= 2,

Y20 =

If

= -4rr cos II,

Y22 =

For I = 3,

Y30

v'4Jt'

=

=

=-

{f

-(3cos2 II - I), 16rr

/lfs hf /if --e 32rr

a ·[g -e'~

8rr

Y21

sin II.

= -

S up .

_e

8rr

sin e cos e,

u« sm ·2 II. T

-(Scos3 II - 3 cos II), 16rr

Y31 = Y32

Yl1

l . 2 sinll(Scos II - I), 64rr

-e'~

OS --e ~ 32rr

2HPsin2 (}cosO,

Y33 = -

/lfS --e 3HPsin3 6. Mrr

From Equations (12.13), (12.18), and (12.34) and the fact that a = 1(1 for some nonnegative integer I, we obtain

+ I)

which gives

d (. dPt) -

I - smll-sin II dll dll

2

-m -211 P Im

sin

+ l(l + I)PIm =

O.

As before, we let u = cos II to obtain 2 d [( l - u 2 dPt] m ] PIm =0. ) - + [1(1+1)---2

-

du

associated Legendre differential equation

du.

1- u

(12.38)

This is called the associated Legendre differential equation. Its solntions, the associated Legendre functions, are given in closed form in Equation (12.33). For m = 0, Equation (12.38) reduces to the Legendre differential equation whose solutions, again given by Equation (12.33) with m = 0, are the Legendre polynomials encountered in Chapter 7. When m = 0, the spherical harmonics become q.>-independent. This corresponds to a physical situation in which there is an explicit azimuthal symmetry. In such cases (when it is obvious that the physical property in question does not depend on q.» a Legendre polynomial, depending only on cos II, will multiply the radial function.

344

12. SEPARATION OF VARIABLES IN SPHERICAL COOROINATES

12.4.1 Expansion of Angular Functions The orthonormality of spherical harmonics can be utilized to expand functions of IJ and rp in terms of them. The fact that these functions are complete will be discussed in a general way in the context of Sturm-Liouville theory. Assuming completeness for now, we write

L~o L~=-I aIm Ylm(IJ, rp) if 1 is not fixed, (12.39)

f(IJ, rp) = {

L~'=-l aIm Ylm(IJ, rp)

if 1 is fixed,

where we have included the case where it is known a priori that f (IJ, rp) has a given fixed 1 value. To find aIm, we multiply both sides by Yl~ (IJ, rp) and integrate over the solid angle. The result, obtained by using the orthonormality relation, is (12.40) where dn es sin IJ dIJ dtp is the element of solid angle. A useful special case of this formula is

alt1 =

ff

dnf(IJ, rp)Y1t(IJ, rp) =

)21/ ff I

dnf(IJ, rp) PI(cos IJ), (12.41)

11:

where we have introduced an extra superscript to emphasize the relation of the expansion coefficients with the function being expanded. Another useful relation is obtained when we let IJ = 0 in Equation (12.39):

f(IJ, rp)19=O =

{

L~O L~=-l almYlm(IJ, rp)19=O if 1 is not fixed, L~=-l aimYlm(IJ , rp)19=O

if 1 is fixed.

From Equations (12.33) and (12.34) one can show that

Therefore,

L~o alt1/ ~t 1

if 1 is not fixed, (12.42)

f(IJ, rp)19=O = [

a(f) 10

/21+1 41l'

if 1 is fixed.

12.4 EIGENVECTORS OF L 2; SPHERICAL HARMONICS

345

z

y

x Figure 12.1 The unit vectors er and er, with their spherical angles and the angle between them.

y

12.4.2 Addition Theorem for Spherical Harmonics

addition theorem for spherical harmonics

An important consequence of the expansion in terms of YZ m is called the addition theorem for spherical harmonics. Consider two unit vectors e, and e" making spherical angles (e, rp) and (e', rp'), respectively, as shown in Figure 12.1. Let y be the angle between the two vectors. The addition theorem states that

L Z

4".* ,, rp,)Yzm(e, rp). Pz(cos y) = ylm(e 2/ + I m=-I

(12.43)

We shall not give a proof of this theorem here and refer the reader to an elegant proof on page 866 which nses the representation theory of groups. The addition theorem is particularly useful in the expansion of the frequently occurring expression 1/lr - r'[. For definiteness we assume [r'] sa r' < [r] es r, Then, introducing t = r' / r, we have 1 1 1 = =-(I+t 2-2tcosy)-1/2. [r - r'l (r 2 + r tl - 2rr' cos y)I/2 r Recalling the generating function for Legendre polynomials from Chapter 7 and using the addition theorem, we get 1

1

-- I'I = r

r

r

00

LtIPI(cosy) [=0

r,l

00

=L [=0

r

4".

Z

1+1 - 2 / 1 L

+

m=-l

Y{;,,(e',rp')Yzm(e,rp)

346

12. SEPARATION OF VARIABLES IN SPHERICAL COOROINATES

It is clear that if r < r', we should expand in terms of the ratio r I r', It is therefore customary to use r < to denote the smaller and r-; to denote the larger of the two

expansion of

radii r and r'. Then the above equation is written as

1/1r - "I inspherical coordinates

(12.44)

This equation is used frequently in the study of Coulomb-like potentials.

12.5 Problems 12.1. By applying the [Xj, Pk]

=

operator [Xj,

pkl to an arbitrary function

fer), show that

i8jk.

12.2. Use the defining relation Li = fijkXjPk to show thatXjPk - XkPj = fijkLi. In both of these expressions a sum over the repeated indices is understood. 12.3. For the angular momentum operator Li = tationrelation [Lj , Lk] = iEjktLI holds.

EijkXj

Pko show that the commu-

12.4. Evaluate aflay and af/az in spherical coordinates and find Ly and Lz in terms of spherical coordinates. 12.5. Obtain an expression for L2 in terms of 8 and
= L+L + L; -

Lz and L2 = L_L+

+ L; + Lz.

12.7. Show that L2, Lx, Ly , and Lz are hermitian operators in the space of squareintegrable functions. 12.8. Verify the following commutation relations:

12.9. Show that L_IY.p) has fJ -I as its eigenvaluefor Lz, and that IY.,P±) cannot be zero. 12.10. Show that if the IYjm) are normalized to mtity, then with proper choice of phase, L IYjm) = ..; j(j + I) - m(m - I) IYj,m-l). 12.11. Derive Equation (12.35).

12.5 PROBLEMS

347

12.12. Starting with Lx and Ly , derive the following expression for L±:

L± = e±i. (±~ + i

ao

cotO~). arp

12.13. Integrate dP/dO -I cotOP = 0 to find P(O). 12.14. Verify the following differential identity:

(:0 + n cot 0) j(O) = sin~ O:e [sin" OJ(O)]. 12.15. Letl = I' and m = m' = 0 in Equation (12.30), and substitnte for YIO from Equation (12.28) to obtain Al = ../(21 + 1)/4".. 12.16. Show that

(-Ile-i(l-kl. d k

LiYI,-/(U, rp) =

c; (I _ u2)(l-k)/2 duk [(I -

12.17. Derive the relations P-m(O) = (_I)m

/

u

2)/].

Y/,-m(O, rp) = (_l)mY/:m(0, rp) and

(I - m)1 . pm(O). (I+m)! /

12.18. Show that I:~=-I IYlm(O, rp)1 2 = (21 + 1)/(4".). Verify this explicitly for 1=landl=2. 12.19. Show that the addition theorem for spherical harmonics can be written as p/(cosy) = PI (cos O)PI (cos 0')

+ 2 I:: /

(I - m)'

m=! (l

.Pt(cos O)Pt (cos 0') cos[m(rp _ rp')].

+m)!

Additional Reading 1. Morse, P. and Feshbach, M. Methods ojTheoretical Physics, McGraw-Hill, 1953. A two-volume classic including a long discussion of the separation of variables in many (sometimes exotic) coordinate systems. 2. The angular momentum eigenvalues and eigenfunctions are discussed in most books on quantnm mechanics. See, e.g., Messiah, A. Quantum Mechanics, volume IT, Wiley, 1966.

13

_

Second-Order Linear Differential Equations

The discussion of Chapter 12 has clearly singled out ODEs, especially those of second order, as objects requiring special attention because most common PDEs of mathematical physics can be separated into ODEs (of second order). This is really an oversimplification of the situation. Many PDEs of physics, both at the fundamental theoretical level (as in the general theory of relativity) and from a practical standpoint (weather forecast) are nonlinear, and the method of the separation of variables does not work. Since no general analytic solutions for such nonlinear systems have been found, we shall confine ourselves to the linear systems, especially those that admit a separated solution. With the exception of the infinite power series, uo systematic method of solving DEs existed during the first halfofthe nineteenth century. The majority of solutions were completely ad hoc and obtained by trial and error, causing frustration and anxiety among mathematicians. It was to overcome this frustration that Sophus Lie, motivated by the newly developed concept of group, took up the systematic study of DEs in the second half of the nineteenth century. This study not only gave a handle on the disarrayed area of DEs, but also gave birth to one of the most beautiful and fundamental branches of mathematical physics, Lie group theory. We shall come back to a thorough treatment of this theory in Parts vn and Vlll. Our main task in this chapter is to study the second-order linear differential equations (SOLDEs). However, to understand SOLDEs, we need some basic understanding of differential equations in general. The next section outlines some essential properties of general DEs. Section 2 is a very brief introduction to firstorder DEs, and the remainder of the chapter deals with SOLDEs.

13.1 GENERAL PROPERTIES OF ODES

13.1

349

General Properties of ODEs

The most general ODE can be expressed as 2y dy d F ( x,y, dx' dx 2 '

···'

d ny) dx n =0,

(13.1)

in which F : R n+2 --> R is a real-valued function of n + 2 real variables. When F depends explicitly and nontrivially on dnyjdx n, Equation (13.1) is called an nth-order ODE. An ODE is said to be linear if the part of the function F that includes y and all its derivatives is linear in y. The most general nth-order linear ODE is po(x)y

homogeneous and inhomogeneous ODEs

dy

dny

+ P!(x)-dx + ... + Pn(X)-d = q(x) xn

for Pn(x) '" 0, (13.2)

where {pilf=o and q are functions of the independent variable x. Equation (13.2) is said to be homogeneous if q = 0; otherwise, it is said to be inhomogeneous and q (x) is called the inhomogeneous term. It is customary, and convenient, to define a linear differential operator L by! L sa po(x)

d

dn

+ P!(x)+ ... + Pn(x)-, dx dx"

Pn(x) '" 0,

(13.3)

and write Equation (13.2) as L[y] =q(x).

(13.4)

A solution of Equation (13.2) or (l3.4) is a single-variable function f : R --> R such that F (x, f(x), f'(x), ... , f(n) (x) = 0, or L[f] = q(x), for all x in the domain of definition of f. The solution of a differential equation may not exist if we put too many restrictions on it. For instance, if we demand that f : R --> R be differentiable too many times, we may not be able to find a solution, as the following example shows. 13.1.1. Example. The mostgeneralsolutioo of dy / dx = Ix Ithat vaoishes at x = 0 is l X2

f( x ) - { ~ - -tx2

ifx>O - ,

if x

:s o.

Thisfunction is continuous and has first derivative f'(x) = [x], which is also continuous at x = O. However, if we demand that its second derivative also be continuous at x = 0, we cannot finda solution, because f"(x)

=

{+I

-I

if x > 0, ifx <0.

1Do not confusethislinear differential operator withthe angular momentum (vector) operator L.

350

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

1m (x)

to exist at x = 0, then we have to expand the notion of a function to include distributions, or generalized functions. II

If we want

. Overrestricting a solution for a differeutial equation results in its absence, but underrestricting it allows multiple solutions. To strike a balance between these two extremes, we agree to make a solution as many times differentiable as plausible and to satisfy certaiu initial conditions. For an nth-order DE such initial conditions are commonly equivalent (but not restricted) to a specification of the function and of its first n - I derivatives. This sort of specification is made feasible by the following theorem. implicit function theorem

13.1.2. Theorem. (implicit functiou theorem) Let G : JRn+l --> lR, given by G(Xl, X2, ... , Xn+l) E JR, have continuous partial derivatives up to the kth order in some neighborhood of a point Po = (rl, ru ... , r,,+I) in JRn+l Let. (8Gj8xn+I)lpo i' O. Then there exists a unique function F : JRn --> JR that is continuously differentiable k times at (some smaller) neighborhood of Po such that Xn+1 = F(Xl, X2, ... ,xn)for all points P = (Xl,X2, ... ,Xn+l) in a neighborhood of Po and G(Xl, X2, ... , Xn, F(XI, X2, ... , x n)) = O. Theorem 13.1.2 simply asserts that under certain (mild) conditions we can "solve" for one of the independent variables in G(XI, X2, ... , x n+! ) = 0 in terms of the others. A proof of this theorem is usually given in advanced calcnlus books. Application of this theorem to Equation (13.1) leads to

2y dny ( dy d d n- 1y) dxn=F x'Y'dx'dx 2 " " ' d xn- 1

'

provided that G satisfies the conditions of the theorem. If we know the solution y = f(x) and its derivatives up to order n - I, we can evaluate its nth derivative using this equation. In addition, we can calculate the derivatives of all orders (assuming they exist) by differentiating this equation. This allows us to expand the solution in a Taylor series. Thus-for solutions that have derivatives of all orders-knowledge of the value of a solution and its first n - I derivatives at a point Xo determines that solution at a neighboring point x. We shall not study the general ODE of Equation (13.1) or even its simpler linear versiou (13.2). We will only briefly study ODEs of the first order in the next section, and then concentrate on linear ODEs of the second order for the rest of this chapter.

13.2 Existence and Uniqueness for First-Order DEs A general first-order DE (FODE) is of the form G(x, y, y') = O. We can find y' (the derivative of y) in terms of a function of x and y if the function G(Xl, x2, X3)

13.2 EXISTENCE AND UNIQUENESS FOR FIRST-ORDER DES

themost general FODE In normal form

explicit solution toa general first-order linear differential equation

351

is differentiable with respect to its third argument and aGjax3 0; O. In that case we have , dy y sa = F(x, y), (13.5) dx which is said to be a normal FODE. If F (x, y) is a linear function of y, then Equation (13.5) becomes a first-order linear DE (FOLDE), which can generally be written as dy (13.6) PI(X)- + po(x)y = q(x). dx It can be shown that the general FOLDE has an explicit solution: (see [Hass 99]) 13.2.1. Theorem. Any first order linear DE oftheform PI (x) y' + Po(x) Y = q (x), in which po, PI, and q are continuous functions in some interval (a, b), has a general solution (13.7)

where C is an arbitrary constantand f1-(x) = _1_ exp PI(X)

[l

X

po(t) dt] , xo PI(t)

(13.8)

where Xo and Xl are arbitrary points in the interval (a, b).

No such explicit solution exists for nonlinear first-order DEs. Nevertheless, it is reassuring to know that a solution of such a DE always exists and under some mild conditions, this solution is unique. We summarizesome of the ideas involved in the proof of the existenceand uniqueuessof the solutionsto FODEs.(Forproofs, see the excellent book by Birkhoff and Rota [Birk 78].) We first state an existeuce theorem due to Peano: Peano existence theorem

13.2.2. Theorem. (Peanoexistencetheorem) Ifthefunction F (x,y) is continuous for the points on and within the rectangle defined by Iy-cl ::s K and Ix -01 ::s N, and if IF(x, y) I ::s M there, then the differential equation y' = F (x, y) has at least one solution, y = f(x), definedfor [x - 01 ::s min(N, K j M) and satisfying the initial condition f(a) == c.

This theorem guaranteesonly the existenceof solutions.Toensure uniqueness, the function F needs to have some additionalproperties. An iroportantproperty is stated in the following definition. Lipschitz condition

13.2.3. Definition. Afunction F(x, y) satisfies a Lipschitz condition in a domain D C IRz iffor some finite constant L (Lipschitzconstant), it satisfies the inequality

IF(x, YI) - F(x, yz)1 s LIYI - yzl for all points (x, YI) and (x, yz) in D.

352

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

uniqueness theorem

13.2.4. Theorem. (uoiqueness theorem) Let f (x) and g(x) be any two solutions ofthe FODE y' = F(x, y) in a domain D, where F satisfies a Lipschitz condition with Lipschitz constant L. Then

If(x) - g(x)1 ::: eLlx-allf(a) - g(a)l. In particular, the FODE has at most one solution curve passing through the point (a, c) E D.

The final conclusion of this theorem is an easy cousequence of the assumed differentiability of F and the requirement f(a) = g(a) = c. The theorem says that if there is a solution y f(x) to the DE y' F(x, y) satisfyiog f(a) c, then it is the solution. The requirements of the Peano existence theorem are too broad to yield solutions that have some nice properties. For iostance, the ioterval of definition of the solutions may depend on their initial values. The followiog example illustrates this poiot.

=

=

=

13.2.5. Example. Consider the DEdyldx = eY • The generalsolntinnof this DE can be obtained by direct integration:

«» dy = Ify

dx

=}

-e- Y = x

+ C.

= b when x = 0, then C = _e- h, and

e-Y

= -x + e-h

=}

y = _In(e- h - x).

Thus, thesolution is defined for -00 < x < e- b , ie., -theinterval of definition of a solution changes with its initialvalue. l1li

To avoid situations illustrated io the example above, one demands not just the continuity of F -as does the Peano existence theorem-but a Lipschitz condition for it. Then one ensuresnot only the existence, but also the uniqueness: local existence and uniqueness theorem

13.2.6. Theorem. (local existence and uniqueness theorem) Suppose that thefunction F(x, y) is defined and continuous in the rectangle Iy - c] :S K, Ix - al ::: N and satisfies a Lipschitz condition there. Let M = max IF(x, y)l in this rectangle. Then the differential equation y' = F(x, y) has a unique solution y = f(x) satisfying f(a) = c and defined on the interval Ix - al :S mio(N, KIM).

13.3 General Properties of SOLDEs The most general SOLDE is

d 2y

P2(X)-2

dx

dy

+ PI (x)-d + po(x)y = x

P3(X).

(13.9)

13.3 GENERAL PROPERTIES OF SOLDES

normal form of a SOLDE

Dividing by P2(X) and writing P for Pl/P2, q for polP2, and r for P3/P2 reduces this to the normal form

d 2y -2

dx

singular points of a SOLDE

regular SOLDE

353

dy

+ P(x)-d + q(x)y = x

(13.10)

rex).

Equation (13.10) is equivalent to (13.9) if P2(X) i' O. The points at which P2(X) vanishes are called the singnlar points of the differential equation. There is a crucial difference between the singular points of linear differential equations and those of nonlinear differential equations. For a nonlinear differential equation such as (x 2_y)y' = x 2+y2, thecnrve y = x 2 is the collection ofsingular points. This makes it impossible to construct solutions y = f (x) that are defined on an interval I = [a, b] of the x-axis because for any x E I, there is a y for which the differential equation is undefined. Linear differential equations do not have this problem, because the coefficients of the derivatives are functions of x only. Therefore, all the singular "curves" are vertical. Thus, we have the following definition. 13.3.1. Definition. The normalform ofa SOWE, Equation (13.10), is regular on an interval [a, b] of the x-axis if p(x), q(x), and rex) are continuous on [a, b]. A solution of a normal SOWE is a twice-differentiable function y = f(x) that satisfies the SOLVE at every point of[a, b].

It is clear that any function that satisfies Equation (13.1O)--{)r Equation (13.9)-must necessarily be twice differentiable, and that is all that is demanded of the solutions. Any higher-order differentiability requirement may be too restrictive, as was pointed out in Example 13.1.1. Most solutions to a normal SOLDE, however, automatically have derivatives of order higher than two. We write Equation (13.9) in the operator form as L[y]=P3,

where

d2 L sa P2dx 2

d

+ PI dx + PO·

(13.11)

It is clear that L is a linear operator because d / dx is linear, as are all powers of it. Thus, for constants a and {J, L[aYI + {JY2] = aL[yJl + {JL[Y2]. In particular, if Yl and Y2 are two solutions of Equation (13.11), then L[YI - Y2] = O. That is, the difference between any two solutions of a SOLDE is a solution of the homogeneous equation obtained by setting P3 = 0. 2 An immediate consequence of the linearity of L is the following: 13.3.2. Lemma. lj'L[u] = rex), L[v] = sex), a and {J are constants, and w = au + {Jy, then L[w] = ar(x) + {Js(x). superposition principle

The proofof this lemma is trivial, but the result describes the fundamental property of linear operators: When r = s = 0, that is, in dealing with homogeneous

2Thisconclusion is, of course, not limited to the SOLDE; it holdsfor all linear DEs.

354

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

equations, the lennna says that any linear combination of solutious of the homogeneous SOLDE (HSOLDE) is also a solution. This is called the superposition principle. Based on physical intuition, we expect to be able to predict the behavior of a physical system if we know the differential equation obeyed by that system, and, equally importantly, the initial data. A prediction is not a prediction unless it is unique.' This expectation for linear equations is bome out in the language of mathematics in the form of an existence theorem and a uniqueness theorem. We consider the latter next. But first, we need a lennna.

s" +

py' + 13.3.3. Lemma. The only solution g(x) ofthe homogeneous equation qy = 0 defined on the interval [a, b] that satisfies g(a) = 0 = g'(a) is the trivial solution g = O. Proof Introduce the nounegative function u(x) es [g(x)]2

+ [g'(x)]2 and differ-

entiate it to get u'(x)

= 2g' g + 2g'g" = 2g'(g + g") = 2g'(g -

pg' - qg)

= _2p(g')2 + 2(1 - q)gg'.

Since (g ± g')2 2: 0, it follows that 21gg'I ::: g2 +

e". Thns,

2(1 - q)gg' ::: 21(1 - q)gg'l = 21(1 - q)1 Igg'l

s

1(1 - q)l(g2 + g'2)

.s (I +

Iql)(g2 + g'2),

and therefore, u'(x)::: lu'(x)1 = 1- 2pg,2 + 2(1- q)gg'l

s 21plg,2 +

(I + Iql)(g2 + gl2)

= [I + Iq(x)l]g2 + [I + Iq(x)1 +2Ip(x)l]gl2. Now let K = I +max[lq(x)1 +2Ip(x)J], where the maximum is taken over [a, b]. Then we obtain u'(x)::: K(g2

+ g'2) =

Ku(x)

V x E [a, b].

Using the resnlt of Problem 13.1 yields u(x) ::: u(a)eK(x-a) for all x E [a, b]. This equation, plus u(a) = 0, as well as the fact that u(x) 2: 0 imply that u(x) = g2(x) + gl2(x) = O. It follows that g(x) = 0 = g'(x) for all x E [a, b]. D uniqueness of solutions fa SOLDE

13.3.4. Theorem. (Uniqueness theorem) Ifp and q are continuous on [a, b], then

3Physical intuition also tells us that if the initial conditions are changed by an infinitesimal amount, then the solutions will be changedinfinitesimally. Thus, the solutions of linear differential equations are said to be continuous functions of the initial conditions. Nonlinear differential equations can have completely different solutions for two initial conditions that are infinitesimally close. Since initialconditions cannot be specifiedwith mathematical precision in practice, nonlinear differential equations leadto unpredictable solutions, or chaos. This subject has receivedmuchattention in recentyears. Foran elementary discussionof chaos see [Hass99, Chapter 15].

13.4 THEWRONSKIAN

355

at most one solution y = f (x) ofEquation (13.10) can satisfy the initial conditions f(a) = Ci and I'(a) = C2, where Ci and C2 are arbitrary constants.

12 be two solutions satisfying the given initial conditions. Then their difference, g es fl - 12, satisfies the bomogeneous equation [witb rex) = 0]. The initial condition that g(x) satisfies is clearly g(a) = a = g'la). By Lemma 13.3.3, g = a or !I = h. 0

Proof Let fl and

Theorem 13.3.4 can be applied to any homogeneous SOLDE to find the latter's most general solution. In particular, let !I (x) and 12 (x) be any two solutions of

y"

+ p(x)y' + q(x)y = 0

(13.12)

defined on the interval [a, b]. Assnme that the two vectors VI = (fl(a), fila»~ and V2 = (h(a), f~(a» in IR2 are linearly independent," Let g(x) be another solution. The vector (g(a), g'(a)) can be written as a linear combination of VI and V2, giving the two equations

g(a) = cdl(a) +c2h(a), g'(a) = Cif{(a) +c2f~(a). Now consider the function u(x) ea g(x) - cdl(x) - c2h(x), which satisfies Equation (13.12) and the initial conditions ural = u'(a) = O. By Lemma 13.3.3, we must have u(x) = 0 or g(x) = Ci!l(X) + c2h(x). We have proved the following: 13.3.5. Theorem. Let

v" + py' + qy

fl and 12 be two solutions ofthe HSOWE

= 0,

where p and q are continuousfunctions defined on the interval [a, b]. If (fl(a), flea»~

and

(h(a), f~(a))

are linearly independent vectors in IR2, then every solution g(x) of this HSOWE is equal to some linear combination g(x) = CI !I (x) +c2h(x) of fl and 12 with constantcoefficients Cl and C2.

13.4 The Wronskian

basis ofsolutions

The two solntions fl(x) and hex) in Theorem 13.3.5 have the property that any other solution g (x) can be expressed as a linear combination of them. We call fl and 12 a basis of solutions of the HSOLDE. To form a basis of solutions, !I

4If theyarenot, thenone mustchoose a different initial pointfor the interval.

356

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

and h must be linearly independent. The linear dependence or independence of a number of functions If; 1i'=1 : [a, b] --> lit is a concept that must hold for all x E [a, b]. Thus, if (a;}i'~1 E lit can be found such that

for some XQ E [a, b], it does not mean that the j's are linearly dependent. Linear dependence requires that the equality hold for all x E [a, b]. In fact, we must write

aIfI

+ a2h + ... + anfn =

0,

where 0 is the zero function. Wronskian defined

13.4.1. Definition. The Wronskian ofany two differentiable functions h(x) is

W(fl,

h: x) = lI(x)f~(x) - h(x)f{(x) = det

(

II (x)

fJ (X)

f{(X») .

h(x)

f~(x)

and

13.4.2. Proposition. The Wronskian ofany two solutions ofEquation (13.12) satisfies

where c is any number in the interval [a, b]. Proof. Differentiating both sides of the definition of Wronskian and substituting from Equation (13.12) yields a FOLDE for W(fl, h: x), which can be easily D solved. The details are left as a problem.

An important consequence of Proposition 13.4.2 is that the Wronskian of any two solutions of Equation (13.12) does not change sign in [a, b]. In particular, if the Wronskian vanishes at one point in [a, b], it vanishes at all points in [a, b]. The real importance of the Wronskian is contained in the following theorem. 13.4.3. Theorem. Two differentiable functions fl and h, which are nonzero in the interval [a', b], are linearly dependent if and only if their Wronskian vanishes.

Proof. If fl and h are linearly dependent, then one is a multiple of the other, and the Wronskian is readily seen to vanish. Conversely, assume that the Wronskian

is zero. Then lI(x)f~(x) - h(x)f{(x)

=0

~ IIdh

and the two functions are linearly dependent.

= hdfl

~ h

= elI D

13.4 THE WRONSKIAN

357

Josef Hoene de Wronski (1778-1853) was hom Josef Hoene, hut he adopted the name Wronski around 1810just after he married, He hadmovedto France andbecome a French citizen in 1800andmovedto Parisin 1810, the same yearhe publishedhis firstmemoiron the foundations of mathematics, which receivedless than favorable reviews from Lacroix andLagrange. His other interests included thedesignof caterpillar vehiclesto competewith therailways. However, theywere nevermanufactured. Wronski was interested mainlyin applying philosophyto mathematics, thephilosophy taking precedence overrigorous mathematical proofs. He criticised Lagrange's use of infinite seriesandintroduced his own ideasfor seriesexpansions of a function. The coefficients in this series are determinants now

known as Wronskians [so named by Thomas Muir (18441934), a Glasgow High School science master who became an authority on determinants by devoting most of his life to writing a five-volume treatise on thehistory of determinants]. Formany years Wronski's workwasdismissedas rubbish. However, a closer examination of the work in more recent timesshows thatalthough someis wrongandhe has anincredibly high opinion of himself andhis ideas,there are also somemathematical insights of great depthandbrilliance hidden within the papers.

13.4.4. Example. Let I: (x) = x and h(x) = Ixl for x E [-1,1]. These two functions are linearly independent in thegiveninterval, becauseal x + a21xI = 0 for all x if andonly if ''1 = ez = O.The Wronskian, on the other hand, vanishes for all x E [-1, + 1]:

dlxl dx dlxl W(ft, fz; x) = x - -Ixl-= x - -Ixl dx dx dx d = x dx

{x

-x

ifx>O if x ~ 0 -

{x

ifx?;O -x if x:::: 0

x =0 if x > 0 -x - (-x) = 0 if x < O.

X -

=

{

Thus, it is possiblefortwo functions to havea vanishing Wronskian without beinglinearly dependent. However, as we showed in the proof of the theorem above, if the functions are differentiable in their interval of definition, then they are linearly dependent if their Wronskian vanishes. II

13.4.5. Example. The Wronskian can be generalized to n functions. The Wronskian of the functions fJ, h. ... , [n is

fJ (x) h(x) :

W(ft,fz, ... ,fn: x) = det

f{(x) f5.(x)

fi"_l)(X») f}"-l) (x)

f/'(x)

fJn-l) (x)

( fn(x)

358

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

If the functions are linearlydependent,then W(ft,!2, ... , in; x) = O. For instance, it is clearthat eX, e- x, andsinhx arelinearly dependent. Thus, we expect W(e X , e- x, sinh x; x)

=

e" e- x ( sinhx

det

e" _e-x cosh x

ex) e- x sinhx

to vanish, as is easilyseen (thefirst and lastcolumns arethesame).

13.4.1 A Second Solution to the HSOLDE If we know one solution to Equation (13.12), say ft, then by differentiating both sides of

= W(x) = W(e)e- J: p(t)dt,

h(x)fi(x) - h(x)f{(x)

dividing the result by f 12 , and noting that the LHS will be the derivative of hih, we can solve for 12 in tenus of ft. The result is

{C + K 1 ---iexp [-1' p(t)dt] dS} , f X

hex) = fleX)

a

l (S)

(13.13)

c

where K '" W (c) is another arbitrary (nonzero) constant; we do not have to know W (x) (this would require knowledge of 12, which we are trying to calculate!) to obtain Wee). In fact, the reader is urged to check directly that hex) satisfies the DE of (13.12) for arbitrary C and K. Whenever possible-and convenientit is customary to set C = 0, because its presence simply gives a term that is proportional to the known solution ft (x). 13.4.6. Example. (a) A solution to the SOLDE y" - k2 y = 0 is ekx. To find a second solution, we let C = 0 and K = I 10Equation (13.13).Since p(x) = 0, we have !2(x)=e

kx

(1

x ds ) I e-2ka 0+ a e2k ' =-2ke-kx+~t!X,

which, ignoring the second term (whichis proportional to thefirst solution), leads directly to thechoice of e- kx as a secondsolution. (h) Thedifferential equationy"+k 2y = ohas sin kx asa solution. WithC = O,a = rr/(2k), andK = I, we get x !2(x) =slokx (o+l ) =-slokxcotksl;/2k = - cos kx .

.~

1r/2k sm ks

(c) For the solotions in part (a), ekx

W(x) = det ( e-kx

kx ke ) -kekx = -2k,

and for those in part (h), W(x) = det (SlokX

coskx

kcoskx ) =-k -ksiokx .

13.4 THE WRONSKIAN

359

BothWronskians are constant. In general, the Wronskian of anytwo linearly independent solutions of y" + q (x) y = 0 is constant. II

Most special functions used in mathematical physics are solutions of SOLDEs. The behavior of these functions at certain special points is determined by the physics of the particular problem. In most situations physical expectation leads to a preference for one particular solution over the other. For example, although there are two linearly independent solutions to the Legendre DE

~ [(I -

Y] dx

x 2) d

dx

+ n(n + I)y =

0,

the solution that is most frequently encountered is the Legendre polynomial Pn (x) discussed in Chapter 7. The other solntion can be obtained by solving the Legendre equation or by using Equation (13.13), as done in the following example.

13.4.7. Example. The Legendreequation can be reexpressed as d 2y 2x dy dx2 - I _ x 2 dx

+

n(n + I) I _ x2 Y = O.

This is an HSOWE with p(x) =

-~2

and

q(x) = n(n

+ I)

I-x 2

I-x

One solution of this HSOLDE is the well-knownLegendre polynomial Pn (x). Using this as ourinputandemploying Equation (13.13), we can generate another set ofsolutions. Let Qn(x) stand for the linearly independent ''partner'' of Pn(x). Then, setting C = 0= c in Equation (13.13) yields x Qn(x) = KPn(x) exp [[' ~ d. " Pn (s) 10 I - t x = KPn(x)lx [~] d. = AnPn(x)l ~. 2 ' a Pn (s} 1-. " (I - . )Pn (.)

1+

-l

+

where An is anarbitrary constant determined by standardization, anda is anarbitrary point +1]. For instance,for n = 0, we have Po = 1, andwe obtain

in the interval [-1,

QO(x)

= Ao

x

1

-d.2

"1-.

= Ao

1 [I2 II+xl I +"1] I-x 2 1-" -In - - - -In 1- -

.

The standard form of Qo(x) is obtained by setting Ao = I and" = 0:

I III-x I

Qo(x) = -In -+x 2

for Ixl < 1.

Similarly, since Pt (x) = x,

II I

x

Ql(X)=AlX

1

2 d. 2 =Ax+Bxln -+x - +C " s (I - . ) I- x

Here standardization is A = 0, B = Ql(X)

I II-+x- I-1. = -xln 2

I-x

!' andC =

for Ixl < 1.

-1. Thus, III

360

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

13.4.2 The General Solution to an ISOLDE Inhomogeneous SOLDEs (ISOLDEs) can be most elegantly discussed iu terms of Green's functions, the subject of Chapter 20, which automatically iucorporate the boundary conditions. However, the most general solution of an ISOLDE, with no boundary specification, can be discussed at this poiut. Let g(x) be a particular solution of L[y]

= y" + py' + qy = r(x)

(13.14)

and let h(x) be any other solution of this equation. Then h(x) - g(x) satisfies Equation (13.12) and can be written as a liuear combiuation of a basis of solutions fl (x) and h (x), leading to the followiug equation:

h(x) = CJfl(x)

method ofvariation ofconstants

+ ezh(x) + g(x).

(13.15)

Thus, if we have apartieular solution of the ISOLDE of Equation (13.14) and two basis solutions of the HSOLDE, then the most general solution of (13.14) can be expressed as the sum of a linear combiuation of the two basis solutions and the particular solution. We know how to find a second solution to the HSOLDE once we know one solution. We now show that knowiug one such solution will also allow us to find a particular solution to the ISOLDE. The method we use is called the variation of constants. This method can also be used to find a second solution to the HSOLDE. Let hand h be the two (known) solutions of the HSOLDE and g(x) the sought-after solution to Equation (13.14). Write g as g(x) = fl (x)v(x) and substitute it in (13.14) to get a SOLDE for v(x):

v"

2 f {) v, =-. r + ( p+-

h

h

This is aftrst order liuear DE iu v', which has a solution of the form

v' = W(x) ft(x)

[c +1 h(t)r(t) dt] W(t) , x

a

where W(x) is the (known) Wronskian of Equation (13.14). Substitutiug

W(x) ft(x)

= h(x)f5.(x) -

h(x)f{(x) ft(x)

=!:... (h) dx

h

iu the above expression for v' and settiug C = 0 (we are interested in a particular solution), we get x dv (h) fl(t)r(t) dt dx dx h a W(t) x x [h(x) h(t)r(t) h(x) h(t)r(t) dt dx h(x) a W(t) fl(x).ax a W(t)

=!:...

1

.».

1

-l _ !:...l

=ft (x)r(x)/W(x)

13.4 THE WRONSKIAN

361

and

r

Ji(t)r(t) dt _ (X h(t)r(t) dt. V(X) = h(x) fl(x) Ja W(t) Ja W(t) This leads to the particular solution x

g(x) = fl(x)v(x) = h(x)

1

fl (t)r(t) W(t) dt - fl(x)

a

1 X

h(t)r(t) W(t) dt.

a

(13.16)

We have just proved the following result. 13.4.8. Proposition. Given a single solution Ji (x) ofthe homogeneous equation corresponding to an ISOLDE, one can use Equation (13.13) to find a second solution h(x) ofthe homogeneous equation and Equation (13.16) to find a particular solution g(x). The most general solution h will then be h(x) = cdl (x)

+ czh(x) + g(x).

13.4.3 Separation and Comparison Theorems The Wronskian can be used to derive some properties ofthe graphs of solutions of HSOLDEs. One such property concerns the relative position of the zeros of two linearly independent solutions of an HSOLDE. theseparation theorem

13.4.9. Theorem. (the separation theorem) The zeros oftwo linearly independent solutions ofan HSOWE occur alternately. Proof Let Ji(x) and h(x) be two independent solutions of Equation (13.12). We have to show that a zero of Ji exists between any two zeros of h. The linear independence of fl and h implies that W (fl, h: x) i' 0 for any x E [a, b]. Let Xi E [a, b] be a zero of h. Then

o i' W(fI,

fz;Xi)

= Ji(xilf~(Xi) -

h(xilf{(xil

= fl(Xi)f~(Xi)'

Thus, [: (Xi) i' 0 and f~(Xi) i' O. Suppose that XI and xz-where X2 > Xl----are two successive zeros of h. Since h is continuous in [a, b] and f~(xI) i' 0, h has to be either increasing [f~(Xl) > 0] or decreasing [f~(XI) < 0] at XI. For h to be zero at xz, the next point, f~ (X2) must have the opposite sign from f~ (Xl) (see Figure 13.1). We proved earlier that the sign of the Wronskian does not change in [a, b] (see Proposition 13.4.2 and comments after it). The above equation then says that fl (XI) and I, (xz) also have opposite signs. The continuity of Ji then implies that Ji must cross the x-axis somewhere between Xl and xz. A similar argument shows that there exists one zero of h between any two zeros of fl. D

362

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

Figure 13.1 If 1 (x I ) > 0 > > 0 > !J (X2).

2

!J (XI)

12(X2). then (assuming thaI the Wronskian is positive)

13.4.10. Example. Two linearly independent sotntions of v" + Y = 0 are sinx and cos x. The separation theorem suggests that the zeros of sin x and cos x must alternate, a fact known from elementary trigonometry: The zeros of cos x occur at odd,multiples of "/2. and those of sin X occur at even multiplesof"/2. ill A second useful result is known as the comparison theorem (for a proof. see [Birk 78. p. 38]). the comparison theorem

13.4.11. Theorem. (the comparison theorem) Let f and g be nontrivial solutions of u" + p (x)u = 0 and v" + q (x) v = O. respectively. where p (x) 2:: q (x) for all X E [a. b]. Then f vanishes at least once between any two zeros of g. unless p = q and f is a constant multiple of g. The form of the differential equations used in the comparison theorem is not restrictive because any HSOLDE can be cast in this form. as the following example shows.

13.4.12. Example. We show that y" u" + S(x)u

+ p(x)y' + q(x)y

=

0 can be cast in the fonn

= Obyanappropriate functional transformation. Define w(x) by y = WU, and

substitutein the HSOLDE to obtain

(u'w

+ w'u)' + p(u'w + w'u) + quw =

wu"

+ (2w' + pw)u' + (qw + pw' + w")u

O.

or = O.

If we demand that the coefficient of u' be zero, we obtain the DE 2w'

solution is w(x) =

cexp[-i

LX p(t)dt].

(B.i?)

+ pw =

0, whose

13.4 THE WRONSKIAN

363

Dividing (13.t7) by this w and substituting for w yields u"

+ S(x)u =

0,

where

Sex) = q

+ pW'- + -w" w

w

12 II = q - 4 P - 'lP .

A useful special case of the comparison theorem is given as the following corollary whose straightforward but instructive proof is left as a problem.

13.4.13. Corollary. If q(x) ::: 0 for all x E [a, bj, then no nontrivial solution of the differential equation v" + q(x)v = 0 can have more than one zero. 13.4.14. Example. It should be clear from the preceding discussion that the oscillations of the solutions of s" + q(x)v = 0 are mostly determined by the sign and magnitude of q(x). For q(x) ::s 0 there is no oscillation; that is, there is no solution that changes sign more than once. Now suppose that q(x) ::: k 2 > 0 for some real k. Then, by Theorem 13.4.11, any solution of v" + q(x)v = 0 must have at least one zero between any two successivezeros of the solution sin kx of ulf + k 2u = O. This means that any solution of v" + q(x)v = ohas a zero in any interval oflength1l"jk if q(x) 2: k 2 > O. Let us apply this to the Bessel DE, y"

2

I , + ( 1- -n ) + -y 2 x

x

y = O.

We can eliminate the y' term by substituting v j.JX for y. 5 This transforms the Bessel DE into.

4n2-1) v" + ( 1 - - - v = O. 2 4x

oscillation ofthe Bessel function of order zero

+

We compare this, for n = 0, with u" U = 0, which has a solution u = sin x, and conclude that each interval of length 7r of the positive x-axis contains at least one zero of any solution of order zero (n = 0) of the Bessel equation. Thus, in particular, the zeroth Bessel function, denoted by Jo(x), has a zero in each interval of length 7r of the x-axis. On the other hand, for 4n2 - I > 0, or n > we have I > [t - (4n2 - l)j4x 2j . This implies that sin x has at least one zero between any two successive zeros of the Bessel functions of order greater than ~. It follows that such a Bessel function can have at most one zero between any two successive zeros of sin x (or in each interval of length 7r on the positive x-axis). 11

!'

13.4.15. Example. Let us apply Corollary 13.4.13 to v" - v = 0 in whichq(x) = -I < O. According to the corollary, the most general solution, cj eX + C2e-x, can have at most one zero. Indeed,

1 c21 ,

x -x = 0 ;::}x=zn1I qe+c2e

q

and this (real) x (if it exists) is the only possible solution, as predicted by the corollary.

SBecause of the square root in the denominator, the range of x will have to be restricted to positive values.

11

364

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

13.5 Adjoint Differential Operators We discussed adjoint operators in detail in the context of finite-dimensional vector spaces in Chapter 2. In particular, the importance of self-adjoint, or hermitian, operators was clearly spelled out by the spectral decomposition theorem of Chapter 4. A consequence of that theorem is the completeness of the eigenvectors of a hermitian operator, the fact that an arbitrary vector can be expressed as a linear combination of the (orthonormal) eigenvectors of a hermitian operator. Self-adjoint differential operators are equally important because their "eigenfunctions" also form complete orthogonal sets, as we shall see later. This section will generalize the concept of the adjoint to the case of a differential operator (of second degree).

13.5.1. Definition. The HSOLVE L[y] '" p2(X)y" + PI (x)y' exact SOLDE

+ po(x)y =

0

(13.18)

is said to be exact if

(13.19) integrating factor for SOLDE

for all f E e 2[a, b] and for some A, B E ella, bJ, An integratingfactorfor L[y] is a function /L(x) such that /L(x)L[y] is exact. If an integrating factor exists, then Equation (13.18) reduces to d, d)A(x)y

+ B(x)y] = 0 =>

A(x)y

,

+ B(x)y = C,

a FOLDE with a constant inhomogeneous term. Even the ISOLDE corresponding to Equation (13.18) can be solved, because

/L(x)L[y]

= /L(x)r(x) => =>

d

dx[A(x)y

,

+ B(x)y] = /L(x)r(x)

A(x)y' + B(x)y =

LX /L(t)r(t)dt,

which is a general FOLDE. Thus, the existence of an integrating factor completely solves a SOLDE. It is therefore important to know whether ornot a SOLDE admits an integrating factor. First let us give a criterion for the exactness of a SOLDE. 13.5,2. Proposition. The SOWE ofEquation (13.18) is exact ifand only if P~ -

pi + Po =

o,

Proof If the SOLDE is exact, then Equation (13.19) holds for all f. implying that P2 = A, PI = A' + B. and PO = B'. It follows that P~ = A", pi = A" + B', and PO = B'. which in tum give P~ - pi + PO = O.

13.5 ADJOINT DIFFERENTIAL OPERATORS

Conversely if Pz - PI + PO = 0, then, substituting PO LHS of Equation (13.18), we obtain

365

= - pz + PI

in the

D

and the DE is exact.

A general SOLDE is clearly not exact. Can we make it exact by multiplying it by an integrating factor as we did with a FOLDE? The following proposition

contains the answer. 13.5.3. Proposition. Afunetion J.L is an integratingfactor ofthe SainE ofEquation (13.18) if and only if it is a solution of the HSOInE M[J.L] es (P2J.L)" - (PIJ.L)' + PoJ.L = O.

(13.20)

Proof This is an immediate conseque,:!ce of Proposition 13.5.2.

D

We can expand Equation (13.20) to obtain the equivalent equation P2J.L"

adjoint of a second-order linear differential operator

+ (2p~ -

PI)J.L'

+ (pz -

PI

+ pO)J.L =

O.

(13.21)

The operator Mgiven by 2

d

M es P2 dx 2

r

+ (2p2 -

d

PI) dx

~

+ (P2" -

,

PI+ PO)

(13.22)

is called the adjoint of the operator L and denoted by Mes Lt. The reason for the use of the word "adjoint" will be made clear below. Proposition 13.5.3 confirms the existence of an integrating factor. However, the latter can be obtained only by solving Equation (13.21), which is at least as difficult as solving the original differential equation! In contrast, the integrating factor for a FOLDE can be obtained by a mere integration [see Equation (13.8)]. Although integrating factors for SOLDEs are not as usefnl as their counterparts for FOLDEs, they can facilitate the study of SOLDEs. Let us first note that the adjoint of the adjoint of a differential operator is the original operator: (Lt)t = L (see Problem 13.11). This suggests that if v is an integrating factor of L[u], then u will be an integrating factor ofM[v] sa Lt[ v]. In particular, multiplying the first one by v and the second one by u and subtracting the results, we obtain [see Equations (13.18) and (13.20)] vL[u] - uM[v] = (VP2)U" - U(P2V)" + (VPI)U' + U(PI v)', which can be simplified to d

,

,

vL[u] - uM[v] = -[p2VU - (P2V) u

dx

+ PIUV].

(13.23)

366

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

Lagrange identities

Integrating this from a to b yields

l

b

(vL[u]- uM[v])dx = [pzvu' - (pzv)'u

+ Pluv]I~.

(13.24)

Equations (13.23) and (13.24) are called the Lagrange identities. Equation (13.24) embodies the reason for calling M the adjoint of L: If we consider u and v as abstract vectors Iu) and Iv), L and M as operators in a Hilbert space with the inner product (ul v) = u*(x)v(x) dx, then Equation (13.24) can be written as

J:

(vi L lu) - (ul M Iv)

= (ul Lt [u)" -

(ul M [u}

= [pzvu' -

(pzv)'u

+ Pluv]I~.

If the RHS is zero, then (ul Lt Iv)* = (ul M Iv) for all ]«) , [u), and since all these operators and functions are real, L t = M. As in the case of finite-dimensional vector spaces, a self-adjoint differential operator merits special consideration. For M[ v] es Lt [v] to be equal to L, we must have [see Equations (13.18) and(13.21)12p~ - PI = PI and P~ - pi + PO = Po· The first equation gives P~ = PI, which also solves the second equation. If this condition holds, then we can write Equation (13.18) as L[y] = PZy" + p~y' + POY, or

d [ d Y] L[y] = dx pz(x) dx

+( Po(x)y = O.

Can we make all SOLDEs self-adjoint? Let us multiply both sides of Equation (13.18) by a function h(x), to be detenntined later. We get the new DE

h(x)pz(x)y" + h(x)PI(x)y' + h(x)po(x)y = 0, which we desire to be self-adjoint. This will be accomplished if we choose h(x) such thathpi (hpz)', or pzh' +h(p~ - PI) 0, which can be readily integrated to give

=

h(x) =

=

~ exp [fX PZ

PI (t) dt] . pz(t)

We have just proved the following: all SOLOEs can be made self-adjoint

13.5.4. Theorem. The SOInE of Equation (13.18) is self-adjoint if and only P~ = PI, in which case the DE has the form

d [ pz(x) dx dY] dx

+ po(x)y =

O.

If it is not self-adjoint, it can be made so by multiplying it through by I h(x) = -exp pz

[fX -PI(t) - d t]. pz(t)

if

13.6 POWER-SERIES SOLUTIONS OF SOLOES

367

13.5.5. Example. (a) The Legendreequationin normalform,

x y" - -2 -x y, +--y=O, I-x l I - x 2

is not self-adjoint. However, we get-aself-adjoint version if we multiplythrough by h(x) I-x 2:

+ Ay =

(I - x 2 )y" - 2xy'

=

0,

or [(I - x 2 )y'1' + Ay = o. (b) Similarly, the normalform of the Besselequation 2

I ( 1 -n- ) y=O y"+-y'+ 2

x

x

is not self-adjoint, but multiplying throughby h(x) = x yields

2

~(XdY) + (x _n ) y = dx. dx x

0,

whichis clearly self-adjoint.

l1li

13.6 Power-Series Solutions of SOLDEs Analysis is one of the richest branches of mathematics, focusing ou the eudless variety of objects we call functions. The simplest kind of functiou is a polynomial, which is obtained by performing the simple algebraic operations of addition and multiplication on the independent variable x. The next in complexity are the trigonometric functions, which are obtained by taking ratios of geometric objects. If we demand a simplistic, intuitive approach to functions, the list ends there. It was only with the advent of derivatives, integrals, and differential equations that a vastly rich variety of functions exploded into existence in the eighteenth and nine-

teenth centuries. Forinstance, e", nonexistent beforetheinvention of calculus, can be thought of as the function that solves dy [dx = y. Although the definition of a function in terms of DEs and integrals seems a bit artificial, for most applications it is the only way to define a function. For instance, theerror function, used in statistics, is.defined as erf(x)

I == r;; y'Jr

IX

e-t 2 dt.

-00

Such a function carmot be expressed in terms of elementary functions. Similarly, functions (of x) such as

1

00

x

sint

-dt, t

368

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

and so on are enconntered frequently in applications. None of these functions can be expressed in terms of other well-known functions. An effective way of studying such functions is to study the differential equations they satisfy. In fact, the majority of functions encountered in mathematical physics obey the HSOLDE of Equation (13.I 8) in which the Pi (x) are elementary functions, mostly ratios of polynomials (of degree at most 2). Of course, to specify functionscompletely,appropriatebonndaryconditions are necessary. For instance, the error function mentioned above satisfies the HSOLDE y" + 2xy' = 0 with the and y'(O) l/..,m. boundary conditions yeO) The natural tendency to resist the idea of a function as a solution of a SOLDE is mostly due to the abstract nature of differential equations. After all, it is easier to imagine constructingfnnctions by simplemultiplications or with simple geometric figures that have been aronnd for centuries. The following beautiful example (see [Birk 78, pp. 85--S7]) should overcome this resistance and convince the skeptic that differential equations contain all the information about a function.

=!

=

13.6.1. Example. We canshow thatthe solutions to y" + Y = 0 have all the properties we expectof sin x and cos x. Let us denote the two linearly independent solutions of this equationbyC(x) andS(x).Tospecifythesefunctiooscompletety, wesetC(O) = S'(O) = t, and C'(O) = S(O) = O. We claim thatthisinformation is enoogh to ideotify C(x) and Sex) as cos x and sinx, respectively. First, let us show that the solutions exist and are well-behaved functions. With C(O) and c' (0) given, the equation y" + y = 0 can generate all derivatives of C (x) at zero: C"(O) =

-C(O) = -I, CII/(O) = -C'(O) = 0, C(4) (0) = -C"(O) = +1, and, in general,

C(n) (0)

=

o {(-I)k

if na iis odd, if a = 2k wherek =

1

0, 1,2, .. ,.

Thns, theTaylor expansion of C(x) is C(x) =

00

2k

k=O

e2k)!

L(-I)k~,

(13.25)

Shnilarty, 00

2k+l

Sex) = "e_l)k_X;-:-:c-; (Zk+I)!'

e13.26)

6

Example illustrates that all intormation about sine and cosine ishidd,m in their differential equation

A simpleratio teston theseriesrepresentation of C(x) yields . ak+t ' e-l)k+l x2(k+I)/e2k+2)!, _x 2 100 - - = too = 1hn =0 k....oo ak k....oo e-l)k x2k/e2k)! k....oo e2k + 2)(2k+ I) ,

whichshowsthat the seriesfor C(x) converges for all valuesof x. Similarly, the seriesfor S(x) is also convergent. Thus, we aredealing with well-defined finite-valued functions. Letus now enumerate and prove some properties of C(x) and S(x). (a) C'ex) = -Sex), We prove this relation by differentiating

c" (x) + C (x)

= 0 and writing the result as

369

13.6 POWER-SERIES SOLUTIONS OF SOLOES

[C' (x) 1" + C' (x) = 0 to make evident the fact that C' (x) is also a solution. Since C' (0) = 0 and [C'eO)]' = C"(O) = -1, and since -Sex) satisfies the same initial conditions, the uniqueness theorem implies that C'(x) = -S(x). Similarly, S'(x) = C(x). (b) C 2(x) +S2(x) = I.

Since the p(x) term is absent from the SOLDE, Proposition 13.4.2implies that theWronskiau of C(x) aud S(x) is constaut. On the otherhaud, W(C, S; x)

= C(x)S' (x)

- C' (x)S(x)

= W(C, S; 0) = C 2(0) (c) S(a + x) = S(a)C(x)

= C 2(x) + S2(x)

+ S2(O) =

I.

+ C(a)S(x).

Theuse oftbe chainruleeasily shows that Sea +x) is a solution of the equation y" + y = O. Thus, it can be written as a linear combination of C(x) and Sex) [which are linearly independent becausetheir Wronskian is nonzeroby (b)]: S(a

+ x) = AS(x) + BC(x).

(\3.27)

This is a functional identity, which for x both sides of Equation (13.27), we get C(a +x) = AS'(x)

+ BC'(x)

= 0 gives

S(a)

= BC(O) = B.lfwe differentiate

= AC(x) - BS(x),

which for x = 0 gives C(a) = A. Substitoting the values of A aud B in Equatiou (13.27) yields the desired identity. A similar argument leads to C(a

+ x)

= C(a)C(x) - S(a)S(x).

(d) Periodicity of C(x) aud S(x). Let xo be the smallest positive real uumber such that S(xo) = C(xo). Then property (b) implies that C(xo) = S(xo) = 1/./2. On the otherhaud, S(xo + x) = S(xO)C(x)

+ C(xo)S(x) =

= C(xo)C(x) - S(xo)S(-x)

C(XO)C(x)

= C(xO -

+ S(xO)S(x)

x).

The third equality follows becaose by Equation (13.26), S(x) is au odd function of x. This is truefor all x; in particular, for x = xo it yields S(2xo) = C(O) = I, aud by property (b), C(2xO) = O. Using property (c) once more, we get S(2xo + x) C(2xo

+ x)

= S(2xo)C(x) + C(2xo)S(x) = C(x), = C(2xo)C(x) - S(2xO)S(x) = -S(x).

Substitoting x = 2xo yields S(4xO) = C(2xo) Continuing in this manner, we can easily obtain S(8xo + x)

= S(x),

C(8xo

+ x)

=

0 aud C(4xo)

=

-S(2xO)

=

-I.

= C(x),

whichprove the periodicity of Sex) and C(x) andshow that their periodis 8xQ. It is even possibleto determine xo. Thisdetermination is left as a problem, buttheresultis xo =

{l/.Ji _d_t_.

10

JJ=t2

A numerical calculation will showthatthis is 11:/4.

370

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

13.6.1 Frobenius Method of Undetermined Coefficients A proper treatment of SOLDEs requires the medium of complex analysis and will be undertakeu in the uext chapter. At this poiut, however, we are seekiug aformal iufiuite series solutiou to the SOLDE

y"

+ p(x)y' + q(x)y =

0,

where p(x) and q(x) are real and analytic. This means that p(x) and q(x) can he represeuted by convergent power series in some interval (a, b). [The interesting case where p(x) and q(x) may have siugularities will be treated in the context of complex solutions.] The general procedure is to write the expansions'' 00

00

p(x) = Lakxk,

(13.28)

q(x) = LbkXk,

k~O

k~O

for the coefficient functions p and q and the solution y, substitute them in the SOLDE, and equate the coefficient of each power of x to zero. For this purpose, we need expansions for derivatives of y:

y'

00

00

k=l

k=O

= LkCkX k- 1 = L(k + l)ck+I Xk, 00

00

v" = L(k + l)kck+IXk-1 = L(k + 2)(k + I)Ck+2 x k k=O

k~l

Thus 00

p(x)y'

=L

00

L

amxm(k + l)Ck+I Xk

= L(k + l)amCk+lX k+m.

k=Om=O

Let k + m

k,m

== n andsumover n. Thenthe other sum,say m, cannot exceed n. Thus, 00

n

p(x)y' = LL(n-m+l)amcn-m+IXn. n=Om=O

Similarly, q(x)y = I:~o I:~~o bmcn_mx n. Substituting these sums and the series for y" in the SOLDE, we obtain

~ { (n + l)(n + 2)Cn+2 +

1;

[en - m

+ l)amCn-m+1 + b~cn-m] } x n =

O.

6Herewe areexpanding about the origin. If suchan expansion is impossible or inconvenient, one can expand about another point,say xo. Onewouldthenreplace all powersof x in all expressions below with powersof x - xo. Theseexpansions assume that p, q, andy haveno singularity atx = O. In general, thisassumption is notvalid,and a different approach, in whichthewhole series is multiplied by a (not necessarily positiveinteger) powerof x, oughtto be taken. Details areprovided in Chapter 14".

13.6 POWER-SERIES SOLUTIONS OF SOLOES

371

For this to be true for all x, the coefficient of each power of x must vanish:

(n + lien

+ 2)c"+2 =

"

- L[(n - m

+ l)amc,,-m+1 + bmc,,_m]

for n 2: 0,

m=O

or ,,-I

n(n + I)C"+l = - L[(n - m)amc,,_m + bmc,,_m_tl

for n 2: 1. (13.29)

m=O

If we know Co and CI (for instance from bonndary conditions), we can uniquely determine c" for n 2: 2 from Eqnation (13.29). This, in tnm, gives a unique powerseries expansion for y, and we have the following theorem. the SOLOE existence theorem

13.6.2. Theorem. (the existence theorem) For any SOLVE of the form s" + p (x) y' + q (x) y = 0 with analytic coefficient functions given by the first two equations of (13.28), there exists a unique power series, given by the third equation of(13.28) that formally satisfies the SOWEfor each choice of co and ct. This theorem merely states the existence of a formal power series and says nothing about its convergence. The following example will demonstrate that convergence is not necessarily guaranteed. 13.6.3. Example. Theformal power-series solution for x 2 s'- y+ x = 0 canbeobtained by letting y = L~o e-x", Then y' = L~o(n + l)cn+lXn, andsubstitution in the DE gives L~O(n + l)Cn+lXn+2 - E~o cnxn + x = 0, or 00

00

L(n+ 1)cn+lXn+2 - Co -qx - LCnxn +x = O. n=O n=2 We see that Co = 0, C1 = 1, and (n + l)Cn+l = cn+2 for n ~ O. Thus, we have the recursion relation nCn = cn+l for n ::: 1 whose unique solution is en = (n - 1)1, which generates the following solution for the DE: Y= x

+ xz + (2!)x 3 + (3!)x 4 + ... + (n -

This series is not convergent for any nonzero x.

t)!x"

+ .... III

As we shall see later, for normal SOLDEs, the power series of y in Equation (13.28) converges to an analytic function. The SOLDE solved in the preceding example is not normal. 13.6.4. Example. As an application of Theorem 13.6.2, tet us cousider the Legendre equation in its normal form x! A Y" - - -2 Z Y +--zy=O. I-x t-x

372

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

For

IxI < I both p and q are analytic, and 00

p(x) = -2x

00

L (x 2)m = L (_2)x 2m+1, m=O

00

q(x)

= ic L

(x 2)m

m=O 00

=

m=O

L

!cx

2m

.

m=O

Thus, the coefficients of Eqoation (13.28) are

am =

o {-2

if m is even.

and

ifmisodd

{~

bm =

if m is even, ifm is odd.

Wewantto substitute for am andbm in Equation (13.29) to find Cn + l . It is convenient + 1, Equation (13.29)-after some algebra-yields to consider two cases: when n is odd and when n is even. For n = 2r r

(2r

L (4r -

+ 1)(2r + 2)C2'+2 =

(13.30)

4m - ic)C2(,-m)'

m=O

With r ---+ r

+ 1, this becomes ,+1

(2r

+ 3)(2r + 4)C2'+4 =

L

(4r m=O

+4 -

4m - ic)C2(,+I-m)

,+1

= (4r

+4 -

= (4r + 4 -

ic)C2(,+I)

+L

(4r

+4 -

4m - ic)C2(,+I_m)

m=l r

ic)c2r+2

+L

(4r - 4m - ic)C2(,_m)

m=O

= (4r

+ 4 - ic)C2,+2 + (2r + 1)(2r + 2)02,+2 + (2r + 3)(2r + 2)]c2,+2,

= [-ic

where in goingfrom the secondequality to the third we changed thedummy index, and in going from the third equality to the fourth we used Eqoation (13.30). Now we let2r+2 ee k to obtain (k + I)(k + 2)Ck+2 = [k(k + I) - iclck, or k(k+ I) -ic

ck+2

= (k + I)(k + 2) Ck

for even k.

It is not difficult to showthat starting with n equation for odd k. Thus, we canwrite n(n+I)-ic cn+2 = (n + I)(n + 2) Cn·

=

2r, thecase of even n, we obtain this same

(13.31)

Forarbitrary co andcj , we obtain twoindependent solutions, one ofwhichhas onlyeven powers of x andtheother onlyoddpowers. Thegeneralized ratiotest(see [Hass99, Chapter 5]) shows thatthe series is divergent for x = ±1 unless A = 1(1 + 1) for some positive

13.6 POWER-SERIES SOLUTIONS OF SOLOES

373

integer l. In that case the infinite series becomes a polynomial, the Legendre polynomial encountered in Chapter 7. Equation (13.31) could have been obtained by substitutiog Equation (13.28) directly intotheLegendre equation. Theroundabout way to (13.31) taken hereshowsthegenerality of Equation (13.29). With specificdifferential equations it is generally better to substitute (13.28) directly. III quantum harmonic oscillator: power series method

13.6.5. Example. Westudied Hermite polynomials in Chapter 7 in the contextof classical orthogonal polynomials. Let us see how theyarisein physics. The one-dimensional time-independent SchrOdinger equation fora particle of mass m in a potential V (x) is 1\2 d 2lj1 - 2m dx2

+ V(x)ljI =

EljI,

where E is the totalenergy of theparticle. For a harmonic oscillator, Vex) = ikx2 ljI" -

m2 w2

~x2lj1

+

== !mw2x2 and

2m

1\2 EljI = O.

Substituting ljI(x) = H(x) exp(-mwx2/21\) and then making the change of variables x = (I/~mw/I\)y yietds

H" -2yH' +J..H

=0

where

2E

J..=--1.

(13.32)

tu»

Thisis theHermite differential equation in normal form. Weassume theexpansion H (y) = L~O cnyn whichyields H'(y)

00

00

n=l

n=O

= Lncnyn-t = L(n + t)cn+1yn, 00

00

H"(y) = Ln(n+l)cn+tyn-l = L(n+l)(n+2)cn+2yn. n=l

n=O

Substituting in Equation (13.32) gives 00

00

L[(n+ I)(n + 2)cn+2 + J..cnJyn - 2 L (n + I)cn+lyn+l = 0, n=O

n=O

or 00

2C2 + J..co

+ L[(n + 2)(n + 3)Cn+3 + J..cn+l -

2(n

+ l)cn+tlY n+l =

n=O

Setting the coefficients of powers of y equal to zero,we obtain C2 =

Cn+3 =

J..

-Zco, 2(n + I) - J.. (n + 2)(n + 3) cn+t

for n

~

0,

O.

374

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS'

or,replacingn withn - 1,

2n-i. c +2 c " - (n+t)(n+2) ",

(13.33)

n2:l.

The ratiotest yields easily thatthe series is convergent for all values of y. Thus, the infinite series whose coefficients obey the recursive relation in Equation (13.33) converges for all y. However, on physical grounds, i.e., the demand that limx--+oo 1fr(x) = 0, the series must be truncated. This happens only if)" = 21 for some integer I (see Probtem 13.20 and [Hass 99, Chapter 13]), and in that case we obtain a polynomial,the Hermite polynomialof order 1. A consequenceof such a truncation is the quantization of harmonic oscillatorenergy: 21

= A=

2E

Iiw - t

=}

E

= (l + !)Iiw.

Two solutions are generated from Equation (13.33), one including only even powers andthe otheronly oddpowers.These areclearlylinearlyindependent. Thus,knowledgeof co and CJ determines the general solution of the HSOLDE of (13.32). III The preceding two examples show how certain special functions used in mathematical physics are obtained in an analytic way, by solving a differential equation. We saw in Chapter 12 how to obtain spherical harmonics and Legendre polynomials by algebraic methods. It is instructive to solve the barmortic oscillator problem using algebraic methods, as the following example demonstrates. quantum harmonic oscillator: algebraic method

13.6.6. Example. TheHamiltonian of a one-dimensional harmonic oscillator is H=

p2

1

2m

2

-+_mo}x 2

where p = -i tid/ dx is themomentum operator. Letusfind theeigenvectors and eigenvalues ofH. We define the operators and Using the commutation relation [x, p] = i1i1, we can show that

and creation and annihilation operators

_ ... t

t ...

H - "wa a + 2,=1.

(13.34)

Furthermore, one can readily show that [H, a] = -liwa,

(13.35)

Let ItfrE) be the eigenvector corresponding to the eigenvalne E: H ItfrE) = E ItfrE}, and note that Equation (13.35) gives Ha ItfrE) = (aH - liwa) /tfrE} = (E - Iiw)a ItfrE) and Hat ItfrE} = (E+Iiw)a t /tfrE). Thus,a /tfrE} is an eigenvectorof H,with eigenvalue E -liw, andat 11/1E} is an eigenvector with eigenvalue E + Iu». That is why at anda arecalledthe raising and lowering (orcreation andannihilation) operators, respectively. We can write

13.6 POWER-SERIES SOLUTIONS OF SOLDES

375

By applying a repeatedly, we obtain states of lower andlowerenergies. But there is a limit to this because His a positiveoperator: It cannot have a negative eigenvalue. Thus, there must exist a ground state, 11{Io}, such that a 11{Io} = O. The energy of this grouod state (or the eigenvaluecorresponding to 11{Io» can be obtained-I

H11{Io}

= (lUvat a + !1Uv) 11{Io) = !1Uv 11{Io) .

Repeated application of theraising operator yieldsbothhigher-level states andeigenvalues. We thus define 11{In} by (13.36)

where en is a normalizing constant. The energy of l1/Jn} is n unitshigher than the ground state's, or

En = (n

t + ~)IUv,

whichis what we obtained in thepreceding example. To find en, we demandorthonormalityfor the 11{In). Takingthe inner product of (13.36) with itself, we can show (see Problem 13.21) that lenl2 = nlc,,_1I2, or len 12 = n!leoI2, which for Icol = 1 and real c" yields en = v'ni.lt follows, then, that (13.37)

quantum harmonic oscillator: connection between algebraic .andanalytic methods

In terms of functions andderivative operators, a Ito} = 0 gives

(J~: x + J2;W :x) 1{Io(x) = 0 with the solution 1{IO(x) = eexp(-mwx 2 /2/i). Normalizing 1{IO(x) gives I

= (1{IoI1{l0) = e2

1

00 -00

2

(mwx )

exp --/i-

dx

lin ) 1/2 = e2 ( mw

Thus,

Wecannow write Equation (13.37) in terms of differential operators: 1{I,,(x) = _1_

v'ni

(~)ti4 lin

Defining a new variable y =

1{1"

(t J wx _

2/i

/i

~)n e-mwx'/(2n)

2mw dx

= .Jmw/hx transforms this equation into

(mW)t/4 _1_ ( _ ~)n e-Y'/2. lin 1],=n Y d -v L.··n:I

7Prom hereon, theunitoperator 1 will notbe shownexplicitly.

Y

376

13. SECONO-OROER LINEAR OIFFERENTIAL EQUATIONS

From this, the relation between Hermite polynomials, and the solutions of the onedimensional harmonic oscillator as given in the previous example, we can obtain a general formula for H n (x). In particular, if we note that (see Problem 13.21)

d) e-Y'/2 = -eY -ed Y

eY '/2 ( y - -

-i

dy

2

dy

and, in general,

eY2 /2 ( y _ _d dy

)n e-Y /2 = (_l)ney 2

ll

2 -e-Y d 2,

dyn

we recover the generalized Rodriguez formula of Chapter 7.

l1li

To end this section, we simply quote the following important theorem (for a proof, see [Birk 78, p. 95]): 13.6.7. Theorem. For any choice of Co and CI, the radius of convergence ofany power series solution y = L~o CkX k for the normal HSOWE

y"

+ p(x)y' + q(x)y =

0

whose coefficients satisfy the recursion relation of (13.29) is at least as large as the smaller ofthe two radii of convergence ofthe two series for p(x) and q(x). 10 particular, if p(x) and q(x) are analytic in an interval around x = 0, then the solution of the normal HSOLDE is also analytic in a neighborhood of x = O.

13.7 SOLDEs with Constant Coefficients The solution to a SOLDE with constant coefficients can always be fonnd in closed form. 10 fact, we can treat an nth-order linear differential equation (NOLDE) with constant coefficients with no extra effort. This brief section outlines the procedure for solving such an equation. For details, the reader is referred to any elementary book on differential equations (see also [Hass 99]). The most general nth-order linear differential equation (NOLDE) with constant coefficients can be written as L[y]

== yCn) + an-1/,,-I) + ... + aIY' + aoy

= r(x).

(13.38)

The corresponding homogeneous NOLDE (HNOLDE) is obtained by setting r(x) = O. Let us consider such a homogeneous case first. The solution to the homogeneous NOLDE

L[y] == yCn)

characteristic polynomial of an HNOLOE

+ an_IY Cn- l) + ... + alY' + aOy =

0

(13.39)

can be found by making the exponential substitotion y = e Ax , which results in the equation L[eAx ] = (A" + a,,_IAn-1 + ... + alA + ao)e AX = O. This equation will hold only if A is a root of the characteristic polynomial

377

13.7 SOLDES WITH CONSTANT COEFFICIENTS

which, by the fundamental theorem of algebra, can be written as (13.40) The Aj are the distinct (complex) roots of peA) and have multiplicity k], 13.7.1. Theorem. Let (Aj J7~1 be the roots ofthe characteristic polynomial ofthe real HNOLDE ofEquation (13.39), and let the respective roots have multiplicities (kj J7=1' Then the functions

are a basis of solutions ofEquation (13.39). When a A is complex, one can write its corresponding solution in terms of trigonometric functions. 13.7.2. Example. An equation that is used in bothmechanics and circuit theory is for a, b > O.

Its characteristic polynomial is pC},,) Al =

!(-a+Ja2 -

4b)

(13.41)

= A2 + a'A + b. whichhastheroots

and

1.2=

i(-a-Ja 2 -

4b).

Wecandistinguish three different possiblemotions depending on therelative sizes of a andb. (a) a 2 > 4b (overdamped): Here we have two distinct simpleroots.The multiplicities are bothone (kl = k2 = 1); therefore, thepower of x forbothsolutions is zeroCrl a 2 - 4b. Then themostgeneral solution is Let y ==

1./

Since a > 2y, this solutionstarts at y = C1 as t -+ 00, yet) -+ o.

+ C2 at t

= rz = 0).

= 0 and continuously decreases; so,

(b) a 2 = 4b (criticallydamped): In this case we have one multipleroot of order 2 (kl = 2); therefore, thepower of x canbe zeroor 1 (rl = 0, 1). Thus, thegeneral solution is yet) = Cjte- at/2 +
This solution starts at y (0) = Co at t = 0, reaches a maximum (or minimum) at t = 2/a - colcr. and subsequently decays (grows)exponentiallyto zero. (e) a 2 < 4b (underdamped): Once more, we havetwo distinct simple roots. The multiplicities arebothone (kl

= k2 =

1); therefore, the power of x for both solutions is zero

378

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

(" = r: = 0). Let '" '" tv'4b - 0 2. Then Al = -0/2 + ico and A2 complex, andthemost general solutionis thusof the form yet)

= e-at j 2 (cr COS(J)t + cz sin cu)

=

= Ai. The roots are

Ae-at j 2 cos(evt + a).

The solution is a harmonic variation witha decayingamplitude A exp(-at /2). Notethatif a = 0, the amplitude does not decay. Thatis why a is called the damping factor (or the damping constant). Theseequations describe either amechanical systemoscillating (withnoexternal force) in a viscous (dissipative) fluid, or an electrical circuit consisting of a resistance R, an inductance L, and a capacitance C. For RLC circuits, a = RJL and b = Ij(LC). Thus, the damping factor depends on therelative magnitudes of R and L. Onthe other hand, the frequency

-

.

depends on all three elements. In particular, for R ~ 2JL/ c the circuit does not oscil-

A physical system whose behavior in the absence ofa driving force is described by a NOLDE will obey an inhomogeneous NOLDE in the presence of the driving force. This driving force is simply the inhomogeneous term of the NOLDE. The best way to solve such an inhomogeneous NOLDE in its most general form is by using Fourier transforms and Green's functions, as we will do in Chapter 20. For the particular, but important, case in which the inhomogeneous term is a product of polynomials and exponentials, the solution can be found in closed form.

13.7.3. Theorem. The NOLDE L[y1= e Ax S(x), where S(x) is a polynomial, has the particular solution eAxq(x), where q(x) is also a polynomial. The degree of q (x) equals that of S(x) unless A = Ai- a root ofthe characteristic polynomial of L, in which case the degree ofq (x) exceeds that of S (x) by ki- the multiplicity of Aj. Once we know the form of the particular solution of the NOLDE, we can find the coefficients in the polynomial of the solution by substitnting in the NOLDE and matching the powers on both sides. 13.7.4. Example. Letus find themostgeneral solutions forthefollowingtwo differential equationssubjectto the bonndarycooditions y(O) = 0 and y' (0) = 1. (a) The first DE we wantto consider is

y" + y

= xe",

(13.42)

The characteristic polynomial is).,2 + I, whose roots areAl = i andA2 = -i. Thus,a basis of solutions is {cos x, sin x}. To find the particular solutionwe note that A(the coefficient of x in the exponential part of the inhomogeneous term) is 1, which is neither of theroots Al and).2- Thus, the particular solutionis of the form q(x)e X , where q(x) = Ax + B is

13.7 SOLDES WITH CONSTANT COEFFICIENTS

of degree I [same degree as that of S(x) = x]. We now substitute u = (Ax Equation (13.42) to obtain the relation

2Axe x

+ (2A + 2B).' =

379

+ B)e X

in

xe",

Matching the coefficients, we have

2A = I

and

2A+2B=O

=}

A=!=-B.

Thus,themost general solution is Y = C1 cosx + C2 sin x

+ 'i
Imposing the given boundary conditions yields 0 Thus, y =

! cos x + sinx + !(x -

=

y(O)

= Cj

-

! and I = y' (0) = C2.

l)e X

is theunique solution. (b) Thenext DE we wantto consider is y" _ y = xe",

(13.43)

I and A2 = -1. A basis of solutions is Here p(A) = A2 - I, and the roots are At {eX,e- X}. To find a particular solotion, we note that S(x) = x and A = I = At. Theorem 13.7.3 then implies thatq(x) must be of degree 2, because At is a simple root, i.e., kt = 1. We therefore try

q(x) = Ax2

+ Bx + C

=}

u = (Ax 2 + Bx + C)e x.

Taking the derivatives and substituting in Eqoation (13.43) yields two equations,

4A= I

and

A+B=O,

1.

whose solution is A = -B = Note thatC is not determined, because Ce" is a solution of the homogeneous DE corresponding to Equation (13.43), so when L is appliedto u, it Another way of looking at the situation is to note that the most eliminates the term C general solution to (13.43) is of the form

ex.

The term Cex could be absorbed in creX. We therefore set C = 0, apply the boundary conditions, andfindtheunique solution III

380

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

13.8 The WKB Method In this section, we treat the somewhat specialized method-due to Wentzel, Kramers, and Brillouin-e-of obtaining an approximate solution to a particular type of second-order DE arising from the Schrodinger equation in one dimension. Suppose we are interested in finding approximate solutions of the DE d2y -2

dx

(13.44)

+q(x)y = 0

in which q varies "slowly" with respect tox in the sense discussed below. If q varies infinitely slowly, i.e., if it is a constant, the solution to Equation (13.44) is simply an imaginary exponential (or trigonometric). So, let us define (x) by y = ei¢(x) and rewrite the DE as

(" - q = O.

(13.45)

Assuming that

=}

= ±

f

(13.46)

Jq(x) dx.

The condition of validity of our assumption is obtained by differentiating (13.46):

1
I

~ ~I « Iql.

It follows from Equation (13.46) and the definition of that 1/y7i is approximately 1/(2Jf) times one "wavelength" of the solution y. Therefore, the approximation is valid if the change in q in one wavelength is small compared to Iq I. The approximation can be improved by inserting the derivative of (13.46) in the DE and solving for a new :

(
12

. , lq

"'q±-2y7i

=}



,

( .zq ' ) t/2

"'± q±-2y7i

or

The two choices give rise to two different solutions, a linear combination of which gives the most general solution. Thus,

y'"

~:(X) {C 1exp [i

f

.jii

dX] + C2 exp [-i f .jii dX]} .

(13.47)

13.8 THE WKB METHOD

381

Equation (13.47) gives an approximate solution to (13.44) in any region in which the condition of validity holds. The method fails if q changes too rapidly or if it is zero at some point of the region. The latter is a serious difficnlty, since we often wish to join a solution in a region in which q (x) > 0 to one in a region in which q (x) < O. There is a general procedure for deriving the so-called connection formulas relating the constants C1 and C2 of the two solutions on either side of the point where q(x) = O. We shall not go into the details of such a derivation, as it is not particularly illuminating.f We simply quote a particular result that is useful in applications. Suppose that q passes through zero at xo, is positive to the right of xo, and satisfies the condition of validity in regions both to the right and to the left of xo. Furthermore, assume that the solution of the DE decreases exponentially to the left of xo. Under such conditions, the solution to the left will be of the form I ~exp[-

-q(x)

l

xO

dX],

(13.48)

Jq(x) dx - ::].

(13.49)

x

J-q(x)

while to the right, we have 2

~q(x) I cos

[l

X

xo

4

A similar procedure gives connection formulas for the case where q is positive on the left and negative on the right of xo. 13.8.1. Example. Consider the Schrodinger equation in onedimension d 2t

dx 2

+

2m 1<2 [E - V(x)lt = 0

where V(x) is a potential well meeting the horizontal line of constant E atx

=

a and x = b,

so that

2m I<

q(x) = Z[E - Vex)]

{><

0 0

if a < x < b, if x < a or x > b.

The solution thatis bounded to the left of a mustbe exponentially decaying. Therefore, in the interval (c, b) the approximate solution, as given by Equation (13.49), is t(x) '"

A 1/4 cos (E - V)

(l

a

X

t~ [E I<

Vex)] dx -

::'4) ,

where A is somearbitrary constant. The solution thatis bounded to therightof b mustalso be exponentially decaying. Hence,the solution for a < x < b is

8Theinterested reader is referred to the bookby Mathews andWalker, pp. 27-37.

382

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

Sincethese twoexpressions give thesamefunction in thesameregion, theymustbe equal. Thus, A

cos

=

B, and, moreimportantly,

(f J~~

[E - V(x)] dx -

~) = cos ( { J~~ [E -

V(x)] dx -

~) ,

or lab JZm[E - V(x)] dx = (n

+ !)rrn.

Thisis essentiallytheBohr-Sommerfeld quantization condition ofpre-1925 quantum mechanics. IIllII

13.8.1 Classical Limit of the Schrodlnger Equation As long as we are approximating solutions of second-order DEs that arise naturally from the Schrodinger eqnation, it is instructive to look at another approximation to the Schr6dinger equation, its classical limit in which the Planck constant goes to zero. The idea is to note that since 1/r(r, t) is a complex function, one can write it as

1/r(r, t) = A(r, t) exp [~s(r, t)] ,

(13.50)

where A(r, t) and S(r, t) are real-valued functions. Snbstituting (13.50) in the Schrodinger equation and separating the real and the imaginary parts yields

(13.51) These two equations are completely equivalent to the Schrodinger equation. The second equation has a direct physical interpretation. Define

p(r, t) == A 2 (r , t) = 11/r(r, t)12

and

VS

J(r, t) sa A 2 (r , t) -

m ___

= pv, (13.52)

=V

multiply the second equation in (13.51) by 2A/m, and note that it then can be written as

ap

-+V·J=O,

at

(13.53)

wltich is the continuity equation for probability. The fact that J is indeed the probability current density is left for Problem 13.30.

13.9 NUMERICAL SOLUTIONS OF DES

The first equation of (13.51) gives an interesting result when 1i, --+ in this limit, the RHS of the equation will be zero, and we get

as

1 2

383

°

because

2

-+-mv +V =0.

at

Taking the gradient of this equation, we obtain

(:t +v'V)mv+VV=O, which is the equation of motion of a classical fluid with velocity field v = V S/ m. We thus have the following: Schriidinger equation describes a classical statistical mixture when Ii -e- O.

13.8.2. Proposition. In the classical limit, the solution of the Schriidinger equation describes a fluid (statistical mixture) of noninteracting classical particles of mass m subject to the potential V(r). The density and the Current density of this fluid are, respectively, the probability density p = I'"12 and the probability current density J of the quantum particle.

13.9 Numerical Solutions of DEs The majority of differential equations encountered in physics do not have known analytic solutions. One therefore resorts to numerical solutions. There are a variety of methods having various degrees of simplicity of use and accuracy. This section considers a few representatives applicable to the solution of ODEs. We make frequent use of techniques developed in Section 2.6. Therefore, the reader is urged to consult that section as needed. Any normal differential equation of nth order,

dnx _ . (n-e l}, --F(x,x, ... ,x ,t), dt n can be reduced to a system of n first-order differential equations by defining Xl = x, Xz

= i, ...

Xn

= x(n-l). This gives the system

We restrict ourselves to a FaDE of the form i = f(x, t) in which f is a wellbehaved function of two real variables. At the end of the section, we briefly outline a technique for solving second-order DEs. Two general types of problems are encountered in applications. An initial value problem (IVP) gives X (t) at an initial time to and asks for the value of X at other times. The second type, the boundary value problem (BVP), applies ouly to differential equations of higher order than first. A second-order BVP specifies the value of x(t) and/or i(t) at one or more points and asks for X or i at other values of t. We shall consider only IVPs.

384

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

13.9.1 Using the Backward Difference Operator Let us consider the IVP i = f(x, r),

(13.54)

x(to) = xo·

The problem is to find {Xk = x(to + kh)}£"=I' given (13.54). Let us begiu by integrating (13.54) between tn aud tn + h: x(tn

+ h) -

x(tn) =

l

to +h

i(t) dt.

to

Chauging the variable of integration to s = (t - tn)/ h aud using the shift operator E introduced in Section 2.6 yields xn+' - Xn

=h

fa'

i(tn + sh) ds

=h

fa\E

S

i(tn)] ds.

(13.55)

Since a typical situation involves calculating x n+' from the values of x(t) aud i(t) at preceding steps, we waut au expression in which the RHS of Equation (13.55) contains such preceding terms. This suggests expressing E in terms of the backward difference operator. It will also be useful to replace the lower limit of integration to - p, where p is a number to be chosen later for convenience. Thus, Equation (13.55) becomes Xn+t = xn- p

+ h [i~ (1 -

[I' ~ 00

= xn -

p

= xn- p

+h

_p

V)-S dS] in

r(-s + l)ds k] . k!r(-s _ k + 1) (-V) X n

+ h (f>iP)V k)

(13.56)

in,

k=O

where (p)

ak

=(-l)kl' -k!

r

P

III

r(-s+l)ds =r(-s - k + 1) k!

-p

s(s+l) .. ·(s+k-l)ds. (13.57)

Keeping the first few terms for p = 0, we obtain the useful formula

Due to the presence of V in Equation (13.58), finding the value of x(t) at tn+1 requires a knowledge of x(t) aud i(t) at points to, tl, ... , tn' Because of this,

13.9 NUMERICAL SOLUTIONS OF DES

formulas of open and closed type

385

Equation (13.58) is called a formula of open type. In contrast, in formulas of closed type, the RHS contains values at tn+1 as well. We can obtain a formula of closed type by changing E' xnto its equivalent form, E,-l xn+ 1. The result is 00

.

Xn+l = x n- p

k. " bCP)V + hLk' Xn+b

(13.59)

k~O

where b CP) es (_l)k

k!

k

1 1

r(-s+2)ds. - k + 2)

_P I'{ -s

Keeping the first few terms for p = 0, we obtain

+ h (1

x +1 "" x n

n

_ V _ V

2

2

12

_ V

3

24

_ 19V

4

720

_ 3V

5

160

_ ... ) X +1

"

,

(13.60)

which involves evaluation at tn +l on the RHS. For p = 1 (p = 3), Equation (13.56) [(13.59)] results in an expansion in powers of V in which the coefficient of V P (V P+2) is zero. Thus, retaining terms up to the (p - l)st [(p + l)st] power of V automatically gives us an accuracy of hl' (h p +2). This is the advantage of using nonzero values of p and the reason we considered such cases. The reason for the use of formulas of the closed type is the smallness of the error involved. All the formulas derived in this section involve powers of V operating on x, or Xn + l . Thismeansthat to findxn+ l , we must know the values of Xk for k :0 n + 1. However, x = f(x, t) or Xk = f(xk, tk) implies that knowledge of Xk requires knowledge of Xk. Therefore, to find Xn+l, we must know not ouly the values of x but also the values of x(t) at tk for k :0 n + 1. In particular, we cannot start with n = 0 because we would get negative indices for x due to the high powers of V. This means that the first few values of Xk must be obtained using a different method. One common method of starting the solution

is to use a Taylor seriesexpansion: h2xo 2 Xk = x (to +kh) =xo +hxok+ -2-k

+ ... ,

(13.61)

where Xo =

Itx«; to),

(8.. xo=

f

ax

I ). + 8f I XQ

Xij,to

-

at

,

XQ,to

For the general case, it is clear that the derivatives required for the RHS of Equation . (13.61) involve very complicated expressions. The following example illustrates the procedure for a specific case.

386

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

13.9.1. Example. Let us solvethe IVP x + x a Taylor series expansion for x by noting that .

2

xQ = -xo -xO'

+ e' x 2 =

0 with x(O) = I. Wecan obtain

xo = x(O) = -XO - 2xOXO - x5,

... .. ".,2 2xQxO.. 4 ' 2 xo=-xO-£...\,oxQxO-xo' Continuing in this way. we can obtain derivatives of all orders. Substituting xQ = 1 and keeping terms up to the fifth order, we obtain

iO = -2,

iQ = 5,

5

x·o = -16,

d4XI

dt 4 ,=0

=65,

d xI = -326. dt 5 ,~O

Substituting these values in a Taylorseries expansionwith h = 0.1 yields Xk = I - 0.2k + 0.025k2 - 0.0027k3

+ (2.7 x

1O-4)k 4 - (2.7 x 1O-5)k5 + ....

Thus, Xt = 0.82254, X2 = 0.68186, and X3 = 0.56741. The corresponding valuesoU can be calculatedusing the DE. Wesimply quote the result: Xt = -1.57026, x2 = -1.24973, X3 = -1.00200. III

Euler's method

Adam's method

Once the starting values are obtained, either a formula of open type or one of closed type is used to find the next x value. Ouly formulas of open type will be discussed here. However, as mentioned earlier, the accuracy of closed-type formulas is better. The price one pays for having x,,+! on the RHS is that using closed-type formulas requires estimating Xn+t. This estimate is then substituted in the RHS, and an improved estimate is found. The process is continued until no further improvement in the estimate is achieved. The use ofopen-type formulas involves simple substitution of the known quantities Xo, xi, ... , Xn on the RHS to obtain x,,+t. The master equation (for p = 0) is (13.58). The number of powers of V that are retained gives rise to different methods. For instance, when no power is retained, the method is called Euler's method, for which we nse xn+t '" x" + hin. A more commouly used method is Adam's method, for which all powers of V up to and including the third are retained. We then have

or, in tenus of values of i,

h

Xn+t '" Xn + 24 (55i n - 59in_t

+ 37i,,_2 -

9i,,_3).

(13.62)

Recall thatik = f(xk, tk). Thus, if we know the values xs, Xn-I, X,,-2, andx,,_3, we can obtain x n+113.9.2. Example. Knowing xo, xI, X2, and X3, we can use Equation (13.62) to calculate X4 for Example 13.9.1: X4 "" X3

+

0.1 24 (55x 3 - 59x2 + 37xl - 9xo) = 0.47793.

13.9 NUMERICAL SOLUTIONS OF DES

With X4 at our disposal, we canevaluate x4 = - X4 -

387

xlet4 , andsubstitute it in

to find X5. and so on.

A crucial fact about such methods is that every value obtaiued is in error by some amount, and using such values to obtain new values propagates the error. Thus, error cau accumulate rapidly aud make approximatious worse at each step. Discussiou of error propagation aud error aualysis-topics that we have not, aud shall not, cover-is commou in the literature (see, for example, [Hild 87, pp. 267268]).

13.9.2 The Runge-Kutta Method The FODE of Equatiou (13.54) leads to a unique Taylor series,

x(to

Runge-Kutta method

+ h) =

h2 xo + hio + 2io + ...

,

where io, io, aud all the rest of the derivatives cau be evaluated by differeutiatiog i = lex, t). Thus, theoretically, the Taylor series gives the solution (for to + h; but to + 2h, to + 3h, aud so on cau be obtaioed similarly). However, in practice, the Taylor series converges slowly, aud the accuracy involved is not high. Thus oue resorts to other methods of solution such as described earlier. Another method, known as the Runge-Kulla method, replaces the Taylor

series (13.63) with

Xn+l = Xn + h [ao!(xn, tn) +

i-»- + J~t

bjh, tn + /Ljh)] , (13.64)

where ao aud {aj, b], /Lj }}=t are constauts choseu such that iftheRHS of (13.64) were expauded in powers of the spacing h, the coefficients of a certaio number of the leading terms would agree with the corresponding expausion coefficieuts of the RHS of (13.63). It is customary to express the b's as linear combinations of preceding values of !; i-I

hb,

= LAirkr, r=O

i = 1,2, ... , p.

388

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

The k; are recursively defined as

= hf(x" + b-h, In + P,rh). Then Equation (13.64) gives Xn+l = Xn + L~=oarkr' The (nontrivial) task now ko = hf(x,,, In),

k;

is to determine the parameters a-, J-Lr, and

Aijo

Carle David Talmo Runge (1856-1927), after returningfrom a six-month vacation in Italy, enrolled at the University of Munich to study literature. However, after six weeks of the course he changed to mathematics and physics. Rungeattendedcourses with Max Planck, andtheybecame close friends. In 1877 both went to Berlin, but Runge turned

to pure mathematics after attending Weierstrass'slectures. His doctoral dissertatioo (1880) dealt with differential geometry. After taking a secondary-school teachers certification test, he returned to Berlin, where he was influenced by Kronecker. Runge then worked on a procedure for the numerical solution of algebraic equations in which the roots were expressed as infinite series of rational functions of the coefficients. In the area of numerical analysis, he is credited with an efficient method of solving differential equations numerically, work he did with Martin Kutta. Runge published little at that stage, but after visiting MitlagMLeffler in Stockholm in

September 1884 he produced a large nnmber of papers in Mittag-Leffler's jonmal Acta mathematica. In 1886. Runge obtained a chair at Hanover and remained there for 18 years. Within a year Runge had moved away from pure mathematics to study the wavelengths of the spectral lines of elements other than hydrogen. He did a great deal of experimental work and published a great quantity of results, including a separation of the spectral lines of helium in two spectral series. In 1904 Klein persuaded Gottingen to offer Runge a chair of applied mathematics, a

post that Rnnge held nnti! he retired in 1925. Runge was always a fit and active man, and on his 70th birthday he entertained his grandchildren by doing handstands. However, a few months later he had a heart attack and

died.

In general, the determination of these constants is extremely tedious. Let us consider the very simple case where p = I, and let A es AOl and IL sa ILl. Then we obtain Xn+l =

x"

+ aoko + alkj,

(13.65)

where ko = hf(xn , In) and k j = hf(xn + Ako, In + p,h). Taylor-expanding kj, a function of two variables, gives?

k: = hf + h 2(lLft + Vfx)

+ h; (1L 2I« + 2Ap,ffxt + A2 f2 fxx) + O(h 4 ) ,

9The symbol O(h ln ) meansthat all termsof order hlll·and higher havebeenneglected.

13.9 NUMERICAL SOLUTIONS OF OES

where It es

anal, etc. Substituting this in the first equation of (13.65), we get

Xn+l = x" w

389

+ h(ao + al)I + h 2a\(J-LIt + Allx) (13.66)

+ h; al(J-L2 Itt + 2AJ-Lllxt + A2 12Ixx) + O(h 4 ) . On the other hand, with

x= x' =

.. df al dx al . x = dt = dl + = xIx 2 Itt + 2IIxt + 1 Ixx + Ix
at

I,

at

+ It

= IIx

+ It,

Equation (13.63) gives

h2 xn+! = Xn + hI + z
h + 6[Itt

cnt in ta ~S.

hs al

+ It) (13.67)

+ 2IIxt + I

2

Ixx

+ Ix
4

).

lfwe demand that (13.66) and (13.67) agree up to the h 2 term (we cannot demand agreement for h 3 or higher because of overspecification), then we mnst have ao + al = 1, alJ-L = ~,aIA = ~' There are ouly three equations for four unknowns. Therefore, there will be an arbitrary parameter f3 in terms of which the unknowns

can be written:

es

ao

= 1- f3,

al = f3,

1 A = 2f3'

a is td

IS

m

i)

Substituting these values in Eqnation (13.65) gives

This formula becomes useful if we let f3 = ~' Then In + hf(2f3) = In + h = In+l, which makes evaluation of the second term in square brackets convenient. For f3 = ~,we have

h X,,+l = x n + "2[f(X n, In) + I(x n + hI, In+I)]

+ O(h 3).

(13.68)

What is nice about this equation is that it needs no starting up! We can plug in the known quantities In, tn+1, and Xn on the RHS and find Xn+I starting with

n = O. However, theresult is not very accurate, and we cannot makeit anymore accurate by demanding agreement for higher powers of h, because, as mentioned earlier, sucha demand overspecifies the unknowns.

390

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

Martin Wilhelm Knlla (1867-1944) lost his parents when he was still a child, and together with his brother went to his uncle in Breslau to go to gymnasium. He attended the University of Bres1au from 1885-1890, and the University of Munich from 1891-1894 concentrating mainly on mathematics, but he was also interested in languages, music, and art. Although he completed the certification for teaching mathematics and physics in 1894, he did not start teaching immediately. Instead, he assisted von Dyckat the Technische Hochschule Miinchen until 1897 (and then again from 1899 to 1903). From 1898 to 1899 he studied at Cambridge, and a year later, he finished his Ph.D. at the University of Munich. In 1902, he completed his habilitation in pure and applied mathematics at the Technische Hochschule Munchen, where he became professor of applied mathematics five years later. In 1909 he accepted an offer from University of Vienna, but a year later he went to the Technische Hochschule Aachen as a professor. From 1912 until his retirement in 1935 he worked at the Technische Hochschule Stuttgart. Kutta's name is well known not only to physicists, applied mathematicians, and engineers, but also to specialists in aerospace science and fluid mechanics. The first group use the Runge-Kutta method, developed at the beginning of the twentieth century to obtain numerical solutions to ordinary differential equations. The second group use the KuttaZhukovskii formula for the theoretical description of the buoyancy of a body immersed in a nonturbulent moving fluid. Kutta's work on the application of conformal mapping to the study of airplane wings was later applied to the flight of birds, and further developed by L. Prandtl in the theory of wings. Kutta obtained the motivation for his first scientific publication from Boltzmann and others (including a historian ofmathematics) when working on the theoretical determination of the heat exchanged between two concentric cylinders kept at constant temperatures. By applying the conformal mapping technique, Kutta managed to obtain numerical values for the heat conductivity of air that agreed well with the experimental values of the time. Three ofKutta's publications dealt with the history ofmathematics, for which he profited greatly because of his knowledge of the Arabic language. One of the most important tasks of applied mathematics is to approximate numerically the initial value problem of ODEs whose solutions cannot be found in closed form. After Euler(1770) had already expressed the basic idea, Runge (1895) and Heun (1900) wrote down the appropriate formulas. Kutta's contribution was to considerably increase the accuracy, and allow for a larger selection of the parameters involved. After accepting a professorship in Stuttgart in 1912, Kutta devoted all his time to teaching. He was very much in demand as a teacher, and it is said that his lectures were so good that even engineering students took an interest in mathematics. (Taken from W. Schulz, "Martin Wilhelm Kutta," Neue Deutsche Biographie 13, Berlin, (1952-) 348-350.)

Formulas that give more accurate results can be obtained by retaining terms beyondp = 1. Thus, for p = 2, if we writexn+ ! = xn I:~~o OI,k" there will be eight unknowns (three 01's, three Aij 's, and two u 's), and the demand for agreement between the Taylor expansion and the expansion of f up to h 3 will yield only six equations. Therefore, there will be two arbitrary parameters whose specification

+

13.9 NUMERICAL SOLUTIONS OF OES

391

results iu various formulas. The details of this kind of algebraic derivation are very messy, so we will merely consider two specific formulas. One suchformula, due to Kutta, is Xn+l = Xn

+ ~(ko + 4kl + k2) + O(h4 ) ,

(13.69)

where

ko = hf(xn , In),

kl = hf(xn

+ !ko, In + !h),

k2 = hf(xn + 2kl - ko, In + h). A second formula, dne to Heun, has the form Xn + ! = Xn

+ !(ko + 3k2) + O(h 4) ,

where

ko = hf(xn, In), ka = hf(xn

+ ~kl

kl = hf(xn + ~kO, In

+ ~h),

- ko, In + ~h).

These two formulas are of about the sarne order of accuracy.

13.9.3. Example. Let us solveIheDE of Example 13.9.1usingIheRunge-Kuua meIhod. WiIh to = 0, xO = I, h = 0.1, and n = 0, Equation (13.69) gives ko = -0.2, kl = -0.17515, k2 = -0.16476. so Ihat XI = 1 + ~(-0.2+4(-0.17515) - 0.16476) = 0.82244.

This Xl, h = 0.1, and tl = to + h = 0.1 yield Ihe following new values: ko = -0.15700, kl = -0.13870, k2 = -0.13040, whicb in turn give

x2 = 0.82244 + ~[-0.15700 - 4(0.13870) - 0.13040] = 0.68207. Wesimilarlyobtain X3 = 0.56964 and X4 = 0.47858. On the oIherhand, solvingIheFODE analyticallygives Ihe exact result x(t) = e-t /(1 + t). Table 13.1 compares the values obtained here, those obtained usingAdam's method, and theexactvalues to fivedecimal places.It is clear that theRunge-Kutta method is more accurate than themethods discussed earlier. II1II The accuracy ofIhe Runge-Kutta meIhod and the fact Ihat it requires no startup procedure make it one of the most popular methods for solving differential equations. The Runge-Kutta method can be made more accurate by using higher values of p, For instance, a formula that is used for p = 3 is Xn+1 = Xn

+ ~(ko + 2kl + 2k2 + k3) + O(h 5 ) ,

where

ko = hf(xn , In), k2 = hf(xn

+ !kl, In + !h)

+ !ko, In + !h), hftx; + k2, In + h).

kl = hf(xn

k3 =

(13.70)

392

13. SECONO-ORDER LINEAR DIFFERENTIAL EQUATIONS

I O.

Analytical 1

Runge--Kutta 1

Adam's method 1

0.1 0.2 0.3 0.4

0.82258 0.68228 0.56986 0.47880

0.82244 0.68207 0.56964 0.47858

0.82254 0.68186 0.56741 0.47793

Table 13.1 Solutions to the differential equation of Example 13.9.1 obtained in three different ways.

13.9.3 Higher-Order Equations Any nth-order differential equation is equivalent to n first-order differential equationsinn+ 1variables. Thus, for instance, the most general SOOE, F(i, X, x, I) = 0, can he reduced to two FOOEs by solving for i to obtain i = G(x, x, z), and defining x = U to get the system of equations

u=

x=u.

G(u, x, I),

These two equations are completely equivalent to the original SOOB. Thus, it is appropriate to discuss numerical solutions of systems of FOOEs in several variables. The discussion here will be limited to systems consisting of two equations. The generalization to several equations is not difficult. Consider the IVP of the following system of equations

x = f(x, u, I),

x(lo)

= xo:

u = g(x, u, I),

u(to)

= uo. (13.71)

Using an obvious generalization of Equation (13.70), we can write X n+l

=

Xn

+ !(ko + 2kl + 2k2 + k3) + O(h s),

Un+l = Un + !(mo + 2ml + 2m2 + m3) + O(h s),

(13.72) where ko

= hf(xn, Un, In),

kl

= hf(xn +

!kO, Un + !mo, In + !h),

k2 = hf(xn + !kl, Un + !mj, In + !h), k3 = hf(xn + ka. Un + m2, In + h),

and mo

= hg(xn, Un, tn ),

mj

= hg(xn +

m2 = hg(xn + !kj, Un + !ml, In + !h), m3 = hg(xn + k2, Un + m2, tn + h).

!ko, Un + !mo, In + !h),

13.9 NUMERICAL SOLUTIONS OF DES

393

These formulas are more general than needed for a SODE, since, as mentioned above, such a SODE is equivalent to the simpler system in which !(x, u, I) sa u. Therefore, Equation (13.72) specializes to

= h(u n + ~mo) = hin + ~hmo, k3 = hin + hm-i,

ko = hu.. k2 =

= hin hin + ~hm"

k,

and

+ ~h(mo + ml + m2) + O(h 5), 5), in+l = in + ~(mo + 2m, + 2m2 + m3) + O(h

Xn+l = Xn + hin

(13.73) where

mo

= hg(xn, in, In),

= hg(xn + ~hin, in + ~mo, In + ~h),

m,

m2 = hg(xn + ~in m3 =

+ ~hmo, in + ~m" In + ~h), hg(xn + hin + ~hm" in + m2, In + h).

13.9.4. Example. The IVP x + x

=

= 0, x(O) = 0, x(O) = I clearly has the analytic

solution xCt) sinr. Nevertheless, let us use Equation (13.73) to illustrate the RungeKutta method and compare the result with the exact solution. For this problem g(x, i, t) = -x. Therefore, we can easily calculate the m's: rna = ":""hxn,

ml = -h(xn

+ !hxn},

+ ~hxn - ~h2Xn), -h[xn + hin - ~h2(Xn + ~hXn)].

m2 = -h(xn m3 =

These lead to the following expressions for x n+1 and xn+1:

x n+ l = xn

+ kin

- ~h2(3Xn

xn+' = in - gh[6Xn

+ kin

+ 3hxn -

- !h2Xn),

h2(xn + !hin)].

Starting with xo = 0 and.to = 1, we can generate xl> X2. and so on by using the last two equations successively. The results for 10 values of x with h = 0.1 are given to five significant figures in Table 13.2. Note that up to x5 there is complete agreement with the exact result. ..

The Runge-Kutta method leuds itself readily to use in computer programs. Because Equation (13.73) does not require any startups, it can be used directly to generate solutions to any IVP involving a SODE. Another, more direct, method of solving higher-order differential equations is to substitute 0 = -(1/ h) 10(1 - V) for the derivative operator in the differential equation, expand in terms of V , and keep an appropriate number of terms. Problem 13.33 illustrates this point for a linear SODE.

394

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

t 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Table 13.'2

Runge-Kutta

0.09983 0.19867 0.29552 0.38942 0.47943 0.56466 0.64425 0.71741 0.78342 0.84161

sint 0.09983 0.19867 0.29552 0.38942 0.47943 0.56464 0.64422 0.71736 0.78333 0.84147

Comparison of the Runge-Kutta andexact solutions to the secondorder DE

of Example 13.9.4.

13.10 Problems 13.1. Let u(x) be a differentiable function satisfying the differential inequality u'(x) ~ Ku(x) for x E [a, b], where K is a constant. Show that u(x) ~ u(a)eK(x-a). Hint: Multiply both sides of the inequality by e-KX, and show that the result can be writteu as the derivative of a nouincreasing fuuction. Then use the fact that a ~ x to get the final result. 13.2. Prove Proposition 13.4.2. 13.3. Let f and g be two differentiable functions that are linearly dependent. Show that their Wronskian vanishes. 13.4. Show that if (Ji, f{) and (!2. f~) are linearly dependent at one point, then Ji and fz are linearly dependent at all x E [a, b]. Here Ji and fz are solutions of the DE of (13.12). Hint: Derive the identity W(fl, f2; X2) = W(Ji, f2; XI) exp

{-1~2 P(t)dt} .

13.5. Show that the solutions to the SOLDE Wronskian.

v" + q(x)y

= 0 have a constant

13.6. Find (in tenus of an integral) GI/(x), the linearly independent "partner" of the Hermite polynomial HI/(x). Specialize this to n = 0, 1. Is it possible to find Go(x) and GI (x) in tenus of elementary functions? 13.7. LetJi, fz,and f3 be any three solutions of y" + py' +qy = O. Show that the (generalized 3 x 3) Wronskian of these solutions is zero. Thus, any three solutions of the HSOLDE are linearly dependent.

13.10 PROBLEMS

13.8. For the HSOLDE v"

+ py' + qy =

11ft. - hI!'

P=

W(fl,

and

h)

395

0, show that

'f "2- 2f'f" q =I 1 1. W(fI,f2)

Thus, knowing two solutions of an HSOLDE allows us to reconstruct the DE.

13.9. Let h, 12, and 13 be three solutions of the third-order linear differential equation s" + P2(X)Y" + PI (x)y' + po(x)y = O. Derive a FODE satisfied by the (generalized 3 x 3) Wronskian of these solutions. 13.10. Prove Corollary 13.4.13. Hint: Consider the solution u = I of the DE u" = 0 and apply Theorem 13.4.11. 13.11. Show that the adjoint of Mgiven in Equation (13.21) is the original L.

13.12. Show that if u(x) and v (x) are solutions of the self-adjointDE (pu')'+qu = 0, then Abel's identity, p(uv' - vu') = constant, holds. 13.13. Reduce each DE to self-adjoint form. (a) x 2y" + xy'

+y =

O.

(b)

s" + y'tan x

= O.

13.14. Reduce the self-adjoint DE (py')' + qy = 0 to u" + S(x)u = 0 by an appropriate change of the dependent variable. What is Sex)? Apply this reduction to the Legendre DE for P"(x), and show that Sex) =

1+ n(n + I) - n(n + l)x 2 . • (l-x2 )2

Now use this result to show that every solution of the Legendre equation has at least (2n + I)/lf zeros on (-I, +1).

13.15. Substitute v = y' / y in the homogeneous SOLDE y"

+ p(x)y' + q(x)y =

0

and:

Riccati equation . (a) Show that it tums into v' + v2 + p(x)v + q(x) = 0, which is a first-order nonlinear equation called the Riccati equation. Would the same substitution work if the DE were inhomogeneous? (h) Show that by an appropriate transformation, the Riccati equation can be directly cast in the form u' + u 2 + Sex) = o.

13.16. For thefunction S(x) defined in Example 13.6.1, let S-I (x) be the inverse, i.e., S-I(S(x» = x. Show that

d

- [S dx

-I

(x)]

=

I vi-x·

"------"2 '

396

13. SECOND-ORDER LINEAR DIFFERENTIAL EQUATIONS

and given that s-t(O) = 0, conclude that x

S-I(X) =

foo

dt ~. 1- t 2

13.17. Define sinh x andcoshx as the solutions of y" = y satisfying the boundary conditions y(O) = 0, y'(O) = I and y(O) = I, y'(O) = 0, respectively. Using Example 13.6.1 as a guide, show that (a) cosh2 x - sinh 2 x = I.

(b) cosh(-x) = coshx.

(c) sinh(-x) = -sinhx.

(d) sinh(a

+ x) = sinha coshx + cosha sinhx.

13.18. (a) Derive Equation (13.30) of Example 13.6.4. (b) Derive Equation (13.31) of Example 13.6.4 by direct substitntion. (c) Let)" = 1(1 + I) in Example 13.6.4 and calculate the Legendre polynomials p/(x) for 1= 0, 1,2,3, subject to the condition p/(l) = I. 13.19. Use Equation (13.33) of Example 13.6.5 to generate the first three Hermite polynomials. Use the normalization

to determine the arbitrary constant. 13.20. The function defined by 00

f(x) = I>nxn,

where

n=O

c

2n -)" C n+ - (n + I)(n + 2) n, 2 -

can be written as f (x) = cog(x) +Cj h (x), where g is even andh is odd in x. Show Xx Xx that f (x) goes to infinity at least as fast as e does, i.e., limx-->oo f (x )e- i' O. Hint: Consider g(x) and h(x) separately and show that 00

g(x) = Lbnx n, n=O

where

b 1n+ - (2n xx

4n -)"

+ 1)(2n + 2) bn' Xx

Then concentrate on the ratio g(x)/e , where g and e are approximated by polynomials of very high degrees. Take the limit of this ratio as x ---> 00, and use Xx recursion relations for g and e • The odd case follows similarly. 13.21. Refer to Example 13.6.6 for this problem. (a) Derive the commutation relation [a, at] = 1. (b) Show that the Hamiltonian can be written as given in Equation (13.34). (c) Derive the commutation relation [a, (at)n] = n(at)n-l.

13.10 PROBLEMS

397

(d) Take the inner product of Equatiou (13.36) with itself aud use (c) to show that len 12 = nlen_11 2 • From this, conclude that Ic,I 2 = n!leoI2. (e) For auy function

(y

f

(y), show that

_~) (eY' /2f) dy

d = _eY' /2 f dy

.

Apply (y - d/dy) repeatedly to both sides of the above equation to obtain

d

n

.

d/' f

) (e Y'/2 f) = (_I)"e Y'/2_. (y _ dy dy" (f) Choose au appropriate f(y) in part (e) aud show that

eY'/2(y _ _d dy

)n e-Y /2 = (-I)'eY -an( e -Y ). 2

2

2

dy"

13.22. Solve Airy's DE, y" + xy = 0, by the power-series method. Show that the radius of couvergence for both independent solutions is infinite. Use the comparisou theorem to show that for x > 0 these solutions have infinitely mauy zeros, but for x < 0 they cau have at most one zero. 13.23. Show that the functions x/' eAX, where r = 0, I, 2, ... , k, are liuearly independeut. Hint: Starting with (D- A)k, apply powers of D- Ato a linear combination of x r eAx for all possible r's. 13.24. Find a basis of real solutions for each DE. (a)

s" + Sy' + 6 =

(b)

O.

d y

(e) -

dx 4

v" +

6y" + 12y' + 8y = O.

d 4y

4

=y.

(d) - 4 =-y.

dx

13.25. Solve the following initial value problems. d 4y (a) - 4 =y,

y(O) = y'(O) = ylll(O) = 0, y"(O) = 1.

dx

d4y

(b) - 4

dx d4y

+

d2y -2

dx

(e) - 4 =0,

dx

=0,

y(O) = y" (0) =

s" (0) =

0, y' (0) = 1.

y(O) = y' (0) = y" (0) = 0,

s" (0) =

2.

13.26. Solve v" -2y' +y = xe" subjectto the initial conditions y(O) = 0, y'(O) = 1.

398

13. SECONO-ORDER LINEAR DIFFERENTIAL EQUATIONS

13.27. Find the general solution of each equation.

(c) y"

+y =

sinx sin 2x.

(e) y" - y = eX sin2x. (g)

(d)

+ 4y = x 2 . v" - y = (1 + e-x )2.

(f)

y(6) _ y(4)

(b) y" - 4y'

(a) y" = xe",

v" -4y' +4 =

eX +xe2x .

(h) y"

+y =

= x 2•

e2x •

13.28. Consider the Euler equation, x n / n) + an_1Xn-1y(n-l)

+ ... + alxy' + aoy

= r(x).

Substitute x = e' and show that such a substitution reduces this to a DE with constant coefficients. In particular, solve x 2y" - 4xy' + 6y = x.

13.29. (a) Show that the substitution (13.50) reduces the Schrodinger equation to (13.51). (b) From the second equation of (13.51), derive the continuity equation for probability. . 13.30. Show that the usual definition of probability current density,

J=Re[l'J*i~ VI'J], reduces to that in Equation (13.52) if we use (13.50).

13.31. Write a computer program that solves the following differential equations by (a) Adam's method [Equation (13.62)] and (b) the Runge-Kutta method [Equation (13.70)]. i = t - x 2, i

= e-

xt,

x(O) = 1 x(O)

i=x 2t2+1,

=

i

1

= t + sinx,

i = sinxt,

x(O) = IT/2 x(O) = 1

x(O) = 1

13.32. Solve the following IVPs numerically, with h = 0.1. Find the first ten values of x. (a)

x + 0.2i 2 + lOx =

(b)x+4x=t 2 , (c)

x+i +x =

0,

(d) tx +i +xt = 0,

(e)

x+i +x

2

= t,

(f) x +xt = 0, (g)

x + sin x = to,

20t,

x(O) = 0,

i(O) = O.

x(O) = I,

i(O) = O.

x(O) = 2,

i(O) = O.

x(O) = I,

i(O) = O.

x(O) = I,

i(O) = O.

x(O) = 0, i(O) = 1. IT x(O) = 2' i(O) = O.

13.10 PROBLEMS

13.33. Substitute dldt = D = -(II h) 1n(1 - V) in the SOLDE i q(t)x = r(t) and expand the log terms to obtain (V 2 + V 3)x n - hp.; (V

+ ~ V 2 )x n + h2qnxn =

399

+ p(t)i +

h 2rn

Since V is of order h, one has to keep one power fewer in the second term, Find an expression for Xn in terms of Xn-l, Xn-2, and Xn-3. validto h 2 .

Additional Reading 1. Birkhoff, G. and Rota, G.-c. Ordinary Differential Equations, 3rd ed., Wiley, 1978. The small size of this book is very deceptive. It is loaded with information. Written by two excellent mathematicians and authors, the book covers all the topics of this chapter and much more in a very clear and lucid style. 2. DeVries, P. A First Course in Computational Physics, Wiley, 1994. The numerical solutions of differential equations are discussed in detail. The approach is slightly different from the one used in this chapter. 3. Hildebrand, F.Introduction to NumericalAnalysis, 2nd ed., Dover, 1987. Our tteatment of numerical solutions of differential equations closely follows that of this reference. 4. Mathews, J. and Walker, R. Mathematical Methods of Physics, 2nd ed., Benjamin, 1970. A good source for WKB approximation.

14

-----'----

_

Complex Analysis of SOLDEs

We have familiarized ourselves with some useful techniques for finding solutions to differential equations. One powerful method that leads to formal solutions is power series. We also stated Theorem 13.6.7 which guarantees the convergence of the solution of the power series within a circle whose size is at least as large as the smallest of the circles of convergence of the coefficient functions. Thus, the convergence of the solution is related to the convergence of the coefficient functions. What about the nature of the convergence, or the analyticity of the solution? Is it related to the analyticity of the coefficient functions? If so, how? Are the singular points of the coefficients also singular points of the solution? Is the nature of the singularities the sarne? This chapter answers some of these questions. Analyticity is best handled in the complex plane. An important reason for this is the property of analytic continuation discussed in Chapter II. The differential equation du [dx = u2 has a solution u = -llx for all x except i = O. Thus, we have to "puncture" the real line by removing x = 0 from it. Then we have two solutions, because the domain of definition of u = -1 I x is not connected on the real line (technically, the definition of a function includes its domain as well as the rule for going from the domain to the range). In addition, if we confine ourselves to the real line, there is no way that we can connect the x > 0 region to the x < 0 region. However, in the complex plane the sarne equation, duif d.; = w 2 , has the complex solution w -liz, which is analytic everywhere except at z O. Puncturing the complex plane does not destroy the connectivity of the region of definition of w. Thus, the solution in the x > 0 region can be analytically continued to the solution in the x < 0 region by going around the origin. The aim of this chapter is to investigate the analytic properties of the solutions of some well known SOLDEs in mathematical physics. We begin with a result from differential equation theory (for a proof, see [Birk 78, p. 223]).

=

=

401

14.1 ANALYTIC PROPERTIES OF COMPLEX DES

continuation principle

14.0.1. Proposition. (continuation principle) The function obtained by analytic continuation of any solution of an analytic differential equation along any path in the complex plane is a solution of the analytic continuation of the differential equation along the same path.

An analytic differential equation is one with analytic coefficient functions. This proposition makes it possible to find a solution in one region of the complex plane and then continue it analytically. The following example shows how the singularities of the coefficient functions affect the behavior of the solution. 14.0.2. Example. LetusconsidertheFODE w'- (y jz)w = Ofory E lit. Thecoefficient functionp(z) = -yjz hasa simplepole at z = O. The solution to theFODEis easilyfound to be w = zY. Thus,depending on whether y is a nonnegative integer, a negative integer -m, or a noninteger, the solution has a regular point, a pole of order m. or a branch point at z = 0, respectively. l1li

This example shows that the singularities of the solution depend on the parameters of the differential equation.

14.1 Analytic Properties of Complex DEs To prepare for discussing the analytic properties of the solutions of SOLDEs, let us consider some general properties of differential equations from a complex analytical point of view.

14.1.1 Complex FOLDEs In the homogeneous FOLDE

dw dz

+ p(z)w =

0,

(14.1)

p(z) is assumed to have only isolated singular points. It follows that p(z) can be expanded about a point zo-which may be a singularity of p(z)-as a Laurent series in some annular region Yj < [z - zo] < r2: 00

p(z)

=

L

an (z - zo)"

where

Yj

zol < rz.

< [z -

n=-oo

The solution to Equation (14.1), as given in Theorem 13.2.1 with q = 0, is

w(z) = exp [- I p(z)dz] = Cexp [-a-II Z

~zzo - ~anl (z 00

= C exp -a_IIn(z - zo) - L [

n=O

an

zo)ndz -

~a-nl (z -

n+l

- - ( z - zo) n+1

zo)-n dz]

a-n-l +L -(z 00

n=l

n

zo)

-n ]

.

402

14. COMPLEX ANALYSIS OF SOLOES

We can write this solution as

w(z) = C(z - zo)"g(z),

(14.2)

where a sa -a-l and g(z) is an analytic single-valued function io the annular region rl < [z - zn] < rz because g(z) is the exponential of an analytic function. For the special case io which p has a simple pole, i.e., when a- n = 0 for all n 2: 2, the second sum io the exponent will be absent, and g will be analytic even at zoo In fact, g(zo) = I, and choosing C = 1, we can write (14.3)

The singularity ofthe coefficient functions of an FOLDE determines the singularity ofthe solution.

Depending on the nature of the siognlarity of p(z) at zo, the solutions given by Equation (14.2) have different classifications. Foriostance, if p(z) has a removable siogularity (if a- n = 0 V n 2: I), the solution is Cg(z), which is analytic. In this case, we say that the FOLDE [Equation (14.1)] has a removable singularity at zoo If p(z) has a simple pole at zo (if a-I f= 0 and a-n = 0 V n 2: 2), then io general, the solution has a branch poiot at zoo In this case we say that the FOLDE has a regular singular point. Fioally, if p(z) has a pole of order m > I, then the solution will have an essential singulatity (see Problem 14.1). In this case the FOLDE is said to have an irregular singular point. To arrive at the solution given by Equation (14.2), we had to solve the FOLDE. Since higher-order differential equations are not as easily solved, it is desirable to obtaio such a solution through other considerations. The followiog example sets the stage for this endeavor. 14.1.1. Example. A FOLDE has a unique solution,to within a multiplicative constant, given by Theorem 13.2.1. Thus, given a solution w(z), any other solutionmustbe of the form Cw(z). Let zu be a singulatity of p(z), and let z - zo = re le . Start at a point z and circle ZQ so that f) --+ e + 21f. Eventhough p(z) mayhavea simple pole at zo, thesolution may have a branch point there. This is clear from the general solution, where a may be a noninteger. Thus, w(z) ea w(zQ + re i (8+ 2:n) may be different from w(z). To discover this branch point-without solving the DE-invoke Proposition 14.0.1 andconclude that w(z) is also a solutionto the FOLDE. Thus, w(z) can be differentfrom w(z) by at most a multiplicative constant: w(z) = Cw(z). Define the complex number ex by C = e'h
+ rei(B+'h<)

a = [ri(B+2"lr w(zo

= (z -

+ re i(8+Z,,)

zo)-ae-Z"iae'h
This argument shows that a solution w(z) of the FOLDE of Equation (14.1) can be written as w(z) = (z - zo)a g(z), where g(z) is single-valued. III

403

14.1 ANALYTIC PROPERTIES OF COMPLEX DES

14.1.2 The Circuit Matrix The method used in Example 14.1.1 can be generalized to obtain a similar result for the NOLDE dnw L[w] = dz n

dn-1w

dw

+ Pn-I(Z) dz n- I + ... + PI(Z) dz + po(z)w =

0

(14.4)

where all the Pi(Z) are analytic in rl < [z - zol < riLet {w j(z)}J~1 be a basis of solutions ofEquation (14.4), andletz-zo = re'", Start at z and analytically continue the functions W j (z) one complete tum to +2IT . Let Wj(z) '" Wj (eo + re ie ) es Wj (zo + rei(9+2"J). Then, by a generalization of Proposition 14.0.1, {Wj(z)}J=1 are not ouly solutions, but they are linearly independent (because they are W j'S evaluated at a different point). Therefore, they also form a basis ofsolutions. On the otherhand, Wj (z) can be expressed as a linearcombinationofthe Wj (z). Thus, Wj(z) = Wj (zo + re i(9+2"») = Lk=1 ajkwk(Z). The matrix A = (a jk), called the circnit matrix of the NOLDE, is invertible, because it transforms one basis into another. Therefore, it has ouly nonzero eigenvalues. We let A be one such eigenvalue, and choose the column vector C, with entries {c, }7~1' to be the corresponding eigenvector of the transpose of A (note that A and At, have the same set of eigenvalues). At least one such eigenvector always exists, because the characteristic polynomial of At has at least one root. Now we let w(z) = LJ=I CjWj(z). Clearly, this w(z) is a solution of (14.4), and

e

circuit matrix

n

w(z) '" w(zo + rei (9+2>tJ) = :~::>jWj(zo + re i(9+2>t J) j=l

n

n

n

= LCj LajkWk(z) = L(At)kjCjWk(Z) = LACkWk(Z) = AW(Z). j=l

k=l

j,k

=

k=l

=

If we define O! by A e2>tia, then w(zo + re i(9+2>tJ) e2"ia w(z). Now we write f(z) ea (z - zo)-aw(z). Following the argument used in Example 14.1.1, we get f(zo + re i(8+2"») = f(z); that is, f(z) is single-valued around ZOo We thus have

the following theorem.

14.1.2. Theorem. Any homogeneous NOWE with analytic coefficient functions in ri < [z - zo 1 < rz admits a solution of the form W(z) = (z - zo)" f(z) where f(z) is single-valued around zo in rl <

lz - eo! < rz-

An isolated singular point zo near which an analytic function w(z) can be written as w(z) = (z - zo)" f(z), where f(z) is single-valued and analytic in the punctured neighborhood of zo. is called a simple branch point of w(z). The arguments leading to Theorem 14.1.2 imply that a solution with a simple branch

404

14. COMPLEX ANALYSIS OF SOLDES

point exists if and only if the vector C whose components appear in w(z) is an eigenvector of At, the transpose of the circuit matrix. Thus, there are as many solutions with simple branch points as there are linearly independent eigenvectors of At

14.2 Complex SOLDEs Let us now consider the SOLDE ui" + p(z)w t ;t- q (z)w = O. Given two linearly independent solutions WI (z) and Wz (z), we form the 2 x 2 matrix A and try to diagonalize it. There are three possible outcomes:

canonical basis ofthe SOLOE

1. The matrix A is diagonalizable, and we can find two eigenvectors, F(z) and G(z), corresponding, respectively, to two distinct eigenvalues, Al and AZ. This means that F(zo + rei(O+z,,)) = Al F(z) and G(zo + rei(O+h)) = AZG(Z). Defining Al = e h i" andAz = e h iP, we get F(z) = (z - zo)" fez) andG(z) = (z-zo)Pg(z),asTheoremI4.1.2suggests. The set {F(z), G(z)} is called a canonical basis of the SOLDE. 2. The matrix Ais diagonalizable, and the two eigenvalues are the same. In this case both F(z) and G(z) have the same constant ex: F(z) =

(z - zo)" fez)

and

G(z) =

(z - zo)"g(z).

3. We cannot find two eigenvectors. This corresponds to the case where A is not diagonalizable. However, we can always find one eigenvector, so A has only one eigenvalue, A. We let WI (z) he the solution of the form (z - zo)" fez), where fez) is single-valued and A = eZ"i". The existence of such a solution is guaranteed by Theorem 14.1.2. Let wz(z) be any other linearly independent solution (Theorem 13.3.5 ensures the existence of such a second solution). Then

and the circuit matrix will be A = (~g), which has eigenvalues A and b. Since A is assumed to have only one eigenvalue (otherwise we would have the first outcome again), we must have b = A. This reduces A to A = (~ ~), where a i" O. The condition a i" 0 is necessary to distinguish this case from the second outcome. Now we analytically continue h(z) sa wZ(Z)/WI(Z) one whole tum around zo, obtaining

14.2 COMPLEX SOLOES

405

It then follows that the function I gl (z)

== h(z) -

a 2:n:iJ.. In(z - zo)

is single-valued in TJ < [z - zn] < rz- If we redefine gl (z) and Wz (z) as (2:n:iJ..ja)gl (z) and (2:n:iJ..ja)wz(z), respectively, we have the following: 14.2.1. Theorem. If p(z) and q(z) are analytic in the annular region rl < [z zol < rz. then the SOWE w" + p(z)w' + q(z)w = 0 admits a basis ofsolutions {WI, wz} in the neighborhood ofthe singular point zo. where either

WI(Z) = (z - zn)" fez), or, in exceptional cases (when the circuit matrix is not diagonalizable], WI (z)

= (z - zo)"fez),

The functions f (z). g(z), and gl (z) are analytic and single-valued in the annular

region. This theorem allows us to factor out the branch point zo from the rest of the solutions. However, even though fez), g(z), and gl (z) are analytic in the annular region rl < Iz - zo I < r2, they may very well have poles of arbitrary orders at zoo Can we also factor out the poles? In general, we cannot; however, under special circumstances, described in the following definition, we can. regular singular point ofa SOLDE defined

14.2.2. Definition. ASOWEoftheform w"+p(z)w'+q(z)w = othat is analytic in 0 < [z - zol < r has a regular singularpoint at zo if p(z) has at worst a simple pole and q(z) has at worst a pole of order 2 there. In a neighborhood of a regular singular point zo, the coefficient functions p(z) and q (z) have the puwer-series expansions 00

a-I p(z) = z -zo q(z) = (

" ak(Z - zo) k , + 'L..., k~O

b-z b_1 )Z + - z-zo z-zo

+~ L...,bk(z k~O

k zo) .

Multiplying both sides of the first equation by z - zo and the second by (z - zo)z and introducing P(z) sa (z - zo)p(z), Q(z) es (z - zo)Zq(z), we obtain 00

P(z) = Lak-I(Z - ZO)k, k=O 1Recallthat In(z - zo) increases by 21ri for eachturn around zoo

00

Q(z) = Lbk-z(z - zo)k. k~O

406

14. COMPLEX ANALYSIS OF SOLDES

It is also convenient to multiply the SOLDE by (z - ZO)2 and write it as (z - ZO)2 w"

+ (z -

zo)P(z)w' + Q(z)w =

o.

(14.5)

Inspired by the discussion leading to Theorem 14.2.1, we write 00

w(z) =

(z - zo)"

L Ck(Z -

Co = I,

ZO)k,

(14.6)

k=O

where we have chosen the arbitrary multiplicative constant in such a way that Co = 1. Substitute this in Equation (14.5), and change the dummy variable-s-so that all sums start at O-to obtain

~ { (n + v)(n + v -

I)Cn

+ ~[(k + v)an-k-I + bn-k-2]Ck }

. (z - zo)"+v = 0,

which results in therecursion relation n

(n

indicial equation, indicial polynomial, characteristic exponents

+ v)(n + v -

I)Cn = - L[(k k=O

+ v)an_k_1 + bn-k-2]Ck·

(14.7)

For n = 0, this leads to what is known as the indicial equation for the exponent

v: ltv) ss v(v -I) +a_Iv

+ b_2 =

O.

(14.8)

The roots of this equation are called the characteristic exponents of zo, and I (v) is called its indicial polynomial. In terms of this polynomial, (14.7) can be expressed as n-I I(n

+ v)Cn =

- L[(k + v)an-k-I k=O

+ bn-k-2]Ck

forn = 1,2, .... (14.9)

Equation (14.8) determines what values of v are possible, and Equation (14.9) gives CI, C2, C3, ... , which in turn determine w(z). Special care must be taken if the indicial polynomial vanishes at n + v for some positive integer n, that is, if n + v, in addition to v, is a root of the indicial polynomial: I (n + v) = 0 = I (v). If uj and V2 are characteristic exponents of the indicial equation and Re(vI) > Re( V2), then a solution for VI always exists. A solution for V2 also exists if VI - v2 # n for any (positive) integer n. In particular, if Zo is an ordinary point [a point at which both p(z) and q(z) are analytic], then only one solution is determined by (14.9). (Why?) The foregoing discussion is summarized in the following:

14.2 COMPLEX SOLDES

407

14.2.3. Theorem. If the dif.ferential equation w" + p(z)w' + q(z)w = 0 has a regular singular point at Z = zo, then at least one power series of the form of (14.6) formally solves the equation. If VI and V2 are the characteristic exponents of zo. then there are two linearly independent formal solutions unless VI - V2 is an integer. 14.2.4. Example. Let us consider some familiar differential equations. (a) The Bessel equation is

w"

I , + ( 1- -a + -w Z

2

z2

)

w·= O.

In this case, the originis a regular singular point, a-I = I, and b-z = _a 2. Thus, the indicia! equation is v(v - 1) + v - a 2 = 0, and its solutions are vI = a and 112 = -a. Therefore, there are two linearly independent solutions to the Bessel equation unless VI - v2 = 2a is aninteger, i.e., unless a is eitheran integeror a half-integer. (b) For the Coulomb potential f (r) = f3Ir, the most general radial eqoation [Equation (12.14)] reducesto

w"

2, + (fJ a) w = + -w -z - -z2 z

O.

The point z = 0 is a regular singular point at which a-I = 2 and b-z = -ct. The indicia! polynomial is lev) = v 2 + v -a withcharacteristic exponents vi = + iJl + 4a and V2 = !oJI + 4a. There are two independentsolutionsunless vI - VZ = oJI + 4a is aninteger. In practice, a = 1(1 + I), where I is some integer; so VI "" vz = 21 + I, and only one solution is obtained. (c) The hypergeometricdifferential equation is

-i

-! -

w"

+y

- (a

+ fJ + I)z w' _

z(1 - z)

~w = O. z(1 - z)

A substantial nwnberof functions in mathematical physicsaresolutions of thisremarkable equation, withappropriate valuesfora, fJ, andy. Theregular singular points- arez = 0 and z = 1. Atz = O,a_1 = y andb_2 = O. The indicial polynomialis I(v) = v(v +y -I), whose roots are VI = 0 and V2 = 1 - y. Unless y is an integer, we have two formal solutions. ..

It is shown in differential equation theory [Birk 78, pp. 40-242] that as long as VI - V2 is not an integer, the series solution of Theorem 14.2.3 is convergent for a neighborhood of zoo What happens when VI - VZ is an integer? First, as a convenience, we translate the coordinate axes so that the point zo coincides with the origin. This will save us some writing, because instead of powers of z - zo, we will have powers of z. Next we let VI = V2 + n with n a positive integer. Then, since it is impossible to encounter any new zero of the indicial polynomial beyond

2Thecoefficientof w neednot havea pole of order 2. Its pole can be of order one as well.

408

14. COMPLEX ANALYSIS OF SOLDES

VI, the recursion relation, Equation (14.9), will be valid for all values ofn, and we obtaina solution: WI (z)

= ZVI f(z) = ZVI

(I + f

Ckl) ,

k=1

which is convergent in the region 0 < Iz I < r for some r > 0 To investigate the nature and the possibility of the second solution, write the recursion relations of Equation (14.9) for the smaller characteristic root V:2: =PII(V2+I)

+ I)CI = 1 (V2 + 2)C2 = I(V2

~

+ b_l) Co -(V:2al + bo)Co -

-(V:2aO

=}

CI = PI,

[(V2

+ I)ao + b_IlCI

=}

C2 sa P2,

(14.10) 1(V:2 + n - I)C,,_I "" p,,-II(V2

+n -

I)Co

1(V:2 + n)C" = I(VI)C" = p"Co

=}

0 = P",

=}

C,,_I = P,,-I,

where in each step, we have used the result of the previous step in which Ck is given as a multiple of Co = 1. Here, the p's are constauts depending (possibly in a very complicated way) on the ak's and bk'S. Theorem 14.2.3 guarantees two power series solutions only when VI - V2 is not an integer. When VI - V2 is an integer, Equation (14.10) shows that ariecessary condition for a second power series solution to exist is that Pn = O. Therefore, when p" i 0, we have to resort to other means of obtaining the second solution. Let us define the second solution as =WI(Z) ,......-"-,

W2(Z) ""

WI (z)h(z)

(14.11)

= ZVl f(z)h(z)

and substitute in the SOLDE to obtain a FOLDE in h', namely, h" + (p + 2w;lwI)h' = O,or, by substituting w;lwI = vi/z-} f' If, the equivalent FOLDE

h" + C~I + 2;' +

p) h'

(14.12)

= O.

14.2.5. Lemma. The coefficient ofh' in Equation (14.12) has a residue ofn

+ 1.

Proof Recall that the residue of a function is the coefficient of Z-I in the Laurent expansion of the function (about Z = 0). Let us denote this residue for the coefficient of h' by A-I. Since f(O) I, the ratio f' If is analytic at z O. Thus, the simple pole at z = 0 comes from the other two terms. Substituting the Laurent expansion of p(z) gives

=

2vI 2vI a-I' - + p = -+-+aO+alz+· ... Z Z Z

=

14.2 COMPLEX SOLOES

409

This shows that A-I = 2vI +0_1. On the other hand, comparing the two versions ofthe indicial polynomial v 2+(a_1 -1)v+b_2 and (v -VI)(V- V2) = v 2 - (VI + 1I2)V + VI V2 gives VI + V2 = -(a-I - I), or 2vI - n = -(a-I - I). Therefore, A_I=2vI+a_l=n+1. 0 14.2.6. Theorem. Suppose that the characteristic exponents of a SOWE with a regular singular point at z = 0 are VI and 1J:2. Consider three cases: 1. VI - Vz is not an integer.

2. V:2

= VI - n where n is a nonnegativeinteger.and Pit, as defined in Equation

(14.10), vanishes.

3.

V2 = VI -

n where n is a nonnegativeinteger, and Pn, as defined in Equation

(14.10), does not vanish.

Then, in the first two cases, there exists a basis ofsolutions {WI, W2} of the form Wi (z) = zV'

(I + f

i

Cki1zk) ,

=

1,2,

k~1

and in the third case, the basis ofsolutions takes the form Wt(z)

= ZV!

(1 + ~akZk),

W2(Z)

= zV'

(1 + ~bkl) +

where the power series are convergent in a neighborhood of z =

CWt(z)lnz,

o.

Proof. The first two cases have been shown before. For the third case, we use Lemma 14.2.5 and write 2vI

2f'

- + -f + p z

n+I = -Z

~

k

+ L... CkZ k~O

,

and the solution for the FOLDE in h' will be [see Equation (14.3) and the discussion preceding it]

h' (z) = z-n-t

(1 + f

bkl) .

k~t

For n = 0, i.e., when the indicial polynomial has a double root, this yields h' (z) = l/z+ I:~1 bkZk- l, or h(z) = Inz+ gl (z), where gl is analytic in a neighborhood ofz = O. Forn oF 0, we have h'(z) = bn/z+ I:i4n bkZk-n- 1 and, by integration,

h(z) = bn ln z +

b L _k_lk;fnk-n 00

n

bk k = bn In z+z -n g2Z, () = bn In z+z -n ~ L...k_nz kim

410

14. COMPLEX ANALYSIS OF SOLOES

where gz is analytic in a neighborhood of z = O. Substituting h in Equation (14.11) and recalling that Vz = VI - n, we obtain the desired results of the theorem. 0

Fuchslan Differential Equations

14.3

In many cases of physical interest, the behavior of the solution of a SOLDE at infinity is important. For instance, bound state solutions of the Schrodinger equation describing the probability amplitndes of particles in quantum mechanics must tend to zero as the distance from the center of the binding force increases. We have seen that the behavior of a solution is determined by the behavior of the coefficient functions. To determine the behavior at infinity, we substitnte z = lit in the SOLDE

dZw

dw

-z + P(z)-d + q(z)w = 0

(14.13)

z

dz

and obtain

-dZvZ + [2- dt

t

1 ] dv -r(t) tZ dt

1 + -s(t)v = 4 t

0,

(14.14)

where v(t) = w(1lt), r(t) = p(llt), ands(t) = q(llt). Clearly, as z --+ 00, t --> O. Thus, we are interested in the behavior of (14.14) at t = O. We assume that both r(t) and s(t) are analytic at t = O. Equation (14.14) shows, however, that the solution v(t) may still have singularities att = 0 because of the extra terms appearing in the coefficient functions. We assume that infinity is a regular singular point of (14.13), by which we mean that t = 0 is a regular singular point of (14.14). Therefore, in the Taylor expansions ofr(t) and s(t), the first (constant) term ofr(t) and the first two terms of s(t) must be zero. Thus, we write 00

r(t)

= alt +azt Z + ... = Lakl, k~l

00

s(t)

= bzt Z +b3t3 + ... = Lbktk. k=Z

By their definitions, these two equations imply that for p(z) and q(z), and for large values of [z], we must have expressions of the form

(14.15)

411

14.3 FUCHSIAN DIFFERENTIAL EQUATlDNS

When infinity is a regular singular point of Equation (14.13), or, equivalently, when the origin is a regular singular point of (14.14), it follows from Theorem 14.2.6 that there exists at least one solution of the form VI (t) = t" (I + L~l ckrk) or, in terms of z, WI (z)

-" (I + ~ Ck)

= z

L..J k

(14.16)

.

k=l Z

Here a is a characteristic exponents at t = 0 of(14.14), whoseindicial polynomial is easily found to be a(a - I) + (2 - al)a + b: = o. Fuchsian DE

Asecond-order Fuchsian DE with two regular singular points leads to uninteresting solutions!

14.3.1. Definition. A homogeneous differential equation with single-valued analytic coefficient functions is called a Fuchsian differential equation (FDE) if it has only regular singular points in the extended complex plane, i.e., the complex plane including the point at infinity. It turns out that a particnlar kind of FDE describes a large class of nonelementary functions encountered in mathematical physics. Therefore, it is instructive to classify various kinds of FDEs. A fact that is used in such a classification is that complex functions whose only singularities in the extended complex plane are poles are rational functions, i.e., ratios of polynomials (see Example 10.2.2). We thus expect FDEs to have only rational functions as coefficients. Consider the case where the equation has at most two regular singnlar points at Zl

and Z2. We introduce anew variable Hz) = Z -

at zt and Z2 are mapped onto the points ~I = ~(ZI) = 0 and ~2 respectively, in the extended ~-plane. Equation (14.13) becomes

d 2u d~2

du

+ (~) d~ + e(~)u =

0,

points

= HZ2) =

00,

(14.17)

where u, <1>, and e are functions of ~ obtained when Z is expressed in terms of in w(z), p(z), and qtz), respectively. From Equation (14.15) and the fact that ~ = 0 is at most a simple pole of (~), we obtain (~) = al/~. Similarly, e(~) = b2/~2. Thus, a SOFDE with two regular singular points is equivaleut to the DE wIt + (al/Z)w' + (b2/Z 2)W = O. Mnltiplying both sides by Z2, we obtain Z2w" + alzw' + b2W = 0, which is the second-order Euler differential equation. A general nth-order Euler differential equation is equivalent to a NOLDE with constant coefficients (see Problem 13.28). Thus, a second order Fuchsian DE (SOFDE) with two regular singular points is eqnivalent to a SOLDE with constant coefficients and produces nothing new. The simplest SOFDE whose solutions may include nonelementary functions is therefore one having three regular singnlar points, at say ZI, Z2, and Z3. By the transformation ~

Asecond-order Fuchsian DE with three regular singular points leads to Interesting solutions!

Zl . The regnlar singular

Z -Z2

~(z) = (z - ZI)(Z3 - Z2) (z - Z2)(Z3 - ZI)

412

14. COMPLEX ANALYSIS OF SOLOES

=

= =

=

we can map Zl, Z2, and Z3 onto ~I 0, ~2 00, and ~3 1. Thus, we assume 0, z I, and z 00. It can be' that the three regular singular points are at Z shown [see Problem (14.8)] that the most geuera! p(z) and q(z) are Al

p(z)

A2

BI

= -z + --1 z-

and

q(z)

=

B2

= 2: + (z- 1)2 z

=

A3

( I) zz-

We thus have the following theorem. Riemann differential equation

14.3.2. Theorem. The most general second order Fuchsian DE with three regular singular points can be transformed into the form /I (AI BI) B2 A3] w=, 0 w + -+- wI+[A2 -+ z Z- I z2 (z - 1)2 z(z - I)

(14.18)

where AI, A2, A3, BI, and B2 are constants. This equation is called the Riemann differential equation.

We can write the Riemaruo DE in terms of pairs of characteristic exponents, (;\,1,1.2), (1-'1,1-'2), and (VI, V2), belonging to the singular points 0, 1, and 00,

respectively. The indicia! equations are easily found to be 2 1. + (AI - 1)1. + A2 = 0,

+ (BI - 1)1-' + B2 = BI)v + A2 + B2 - A3 =

0,

1-'2

v

2

+ (I -

Al -

O.

By writing the indicia! equations as (A - 1.1)(1. - 1.2) = 0, and so forth and comparing coefficients, we can find the following relations: Al =

I - Al - 1.2,

BI =

I - 1-'1 - 1-'2,

A2 = 1.11.2, B2 = 1-'11-'2,

AI+BI =VI+V2+1,

A2

+ B2 -

A 3 = VI V2.

These equations lead easily to the Riemann identity Al

+ 1.2 + 1-'1 + 1-'2 + VI + V2 =

1.

(14.19)

Substituting these results in (14.18) gives the following result. 14.3.3. Theorem. A second order Fuchsian DE with three regular singuiarpoints in the extended complex plane is equivalent to the Riemann DE, wI! + ( 1 - 1.1 - 1.2 Z

1.11.2 + [ --+ Z2

+ 1- "1- r-"2) w' t'"

z-I 1-'11-'2 1.11.2 -1-'11-'2] + VIV2 - z(z w-O (z - 1)2 - I) - ,

(14.20)

which is uniquely determined by the pairs of characteristic exponents at each singularpoint. The characteristic exponents satisfy the Riemann identity, Equation (14.19).

14.4 THE HYPERGEOMETRIC FUNCTION

413

The uniqueness of the Riemann DE allows us to derive identities for solutions and reduce the independent parameters of Equation (14.20) from five to three. We first note that if w(z) is a solution of the Riemann DE corresponding to (AI, A2), (J.'I, J.'2), and (VI, 112), then the function u(z) = z},(z - I)JLw(z)

has branch points at z = 0, 1,00 [because w(z) does]; therefore, it is a solution of the Riemann DE. Its pairs of characteristic exponents are (see Problem 14.10) (AI +A,A2+A),

In particular, if we let A = -AI and J.' = - J.'1, then the pairs reduce to (0, J.'2 - J.'I),

Defining a sa VI + Al + J.L!, fJ es V2 + Al + J.'I, and y (14.19), we can write the pairs as (0, y - a - fJ),

(O,I-y),

== I -

A2 + AI, and using

(a, fJ),

which yield the third version of the Riemann DE w" +

(1:'. + z

hypergeometric DE

1- Y + a + fJ) w' + afJ w = O. z- I z(z - 1)

This important equation is commonly written in the equivalent form z(1 - z)w" + [y - (1 + a + fJ)z]w' - afJw = 0

(14.21)

and is called the hypergeometric differential equation (HGDE). We will study this equation next.

14.4 The Hypergeometric Function

hypergeomelric function

The two characteristic exponents of Equation (14.21) at z = 0 are 0 and I - y. It follows from Theorem 14.2.6 that there exists an analytic solution (corresponding to the characteristic exponent 0) at z = O. Let us denote this solution, the hypergeometric function, by F(a, fJ; y; z) and write 00

F(a, fJ; y; z) = I>kZk

where

ao = I.

k~O

Substituting in the DE, we obtain the recurrence relation ak+I

(a = (k

+ k)(fJ + k) + I)(y + k) ak

for k 2: O.

414

14. COMPLEX ANALYSIS OF SOLOES

hypergeometric series

Thesecoefficients canbe determined successivelyif y is neither zeronoranegative integer: . . _ ~ a(a + 1)··· (a + k - 1)f:l(f:l + 1)· .. (f:l + k - 1) k F(a,f:l,y,z)-l+t;r k!y(y+1) ... (y+k-1) z

r(y)

= r(a)r(f:l)

~ r(a + k)r(f:l + k) k

6

r(k + l)r(y + k) z .

(14.22)

The series in (14.22) is called the hypergeometric series, because it is the generalization of F(l, f:l; f:l; z), which is simply the geometric series. We note immediately from (14.22) that

14.4.1. Box. The hypergeometric series becomes a polynomial or f:l is a negative integer.

if either a

Tbisis because for k < lal (ork < If:ll)bothr(a+k) [orr(f:l+k)]and I'(o) [or r(f:l)] have poles that cancel each other. However, r(a + k) [or r(f:l + k)] becomes finitefor k > lal (or k > If:ll), and the pole in I'(o) [or r(f:l)] makes the denominator infinite. Therefore, all terms of the series (14.22) beyond k = la I (or k = If:lll will be zero. Many of the properties of the hypergeometric function can be obtained directly from the HGDE, Equation (14.21). For instance, differentiating the HGDE and letting v = w', we obtain z(l - z)v" + [y + 1 - (a + f:l + 3)zlv' - (a + 1)(f:l+ l)v = 0, which shows that F'(a, f:l; y; z) = CF(a + 1, f:l + 1; y + 1; z), The constant C can be determined by differentiating Equation (14.22), setting z = 0 in the result.' and noting that F(a + I, f:l + 1; y + 1; 0) = 1. Then we obtain af:l F'(a,f:l; y; z) = -F(a + I, f:l + 1; y + 1; z). y

(14.23)

Now assume that y '" 1, and make the substitution w = Zl-Yu in the HGDE to obtain" z(l-z)u" +[Yl - (al +f:ll + l)z]u' -alf:llu = 0, where al = a -Y + 1, f:ll = f:l- Y + I, and Yl = 2 - y. Thus, u = F(a - y + I, f:l- Y + 1; 2 - y; z), 3Notethat thehypergeometric function evaluates to 1 at Z = 0 regardless of its parameters. 4In thefollowingdiscussion, aI, Ih, andYl will represent theparameters of thenew DE satisfied by thenew function defined in terms of the old.

14.4 THE HYPERGEOMETRIC FUNCTION

415

and u is therefore analytic at z = D. This leads to an interesting resnIt. Provided that y is not an integer, the two functions Wt(z)

== F(a, fJ; y; z),

W2(Z) == zt-y F(a - y

+ 1, fJ -

Y + I; 2 - y; z) (14.24)

form a canonical basis of solutions to the HGDE at Z = D. This follows from Theorem 14.2.6 and the fact that (D, I - y) are a pair of (different) characteristic exponents at z = D.

Johann Carl Friedrich Gauss (1777-1855)was thegreatestof all mathematiciansandperhaps the most richly gifted genius of whom there is any record. He was born in the city of Brunswick in northern Germany, His exceptional skill with numbers was clear at a very early age, and in later life he joked that he knew how to count before he could talk. It is said that Goethe wrote

and directedlittle plays for a puppet theaterwhen he was 6 and that Mozart composed his first childish minuets when he was

5, but Gauss corrected an error in his father's payroll accounts at the age of 3. At the age of seven, when he started elementary school, his teacher was amazed when Gauss summed the integers from 1 to 100 instantly by spotting that the som was

50 pairs of numbers each pair summing to 1Oi. His long professional life is so filled with accomplishments that it is impossible to give a full account of them in the short space available here. All we can do is simply give a chronology of his almost uncountable discoveries. 1792-1794: Gauss reads the works of Newton, Euler, and Lagrange; discovers the prime number theorem (at the age of 14 or 15); invents the method of least squares; conceives the Gaussian law of distribution in the theory of probability. 1795: (only 18 years old!) Proves tbataregularpolygon withn sides is constructible (by ruler and compass) if and only if n is the product of a power of2 and distinct prime numbers of the 2k

form Pk = 2 + 1, and completely solves the 200D-year old problem of ruler-and-compass construction of regular polygons. He also discovers the law of quadratic reciprocity. 1799: Proves the fundamental theorem ef algebra in his doctoral dissertation using the then-mysterious complex numbers with complete confidence. 1801: Gauss publishes his Dtsquistitones Arithmeticae in which he creates the modem rigorous approach to mathematics; predicts the exact location of the asteroid Ceres. 1807: Becomes professor of astronomy and the director of the new observatory at Gottingen, 1809: Publishes his second book, Theoria motus corporum coelestium, a major two-volume treatise on the motion of celestial bodies and the bible of planetary astronomers for the next

100 years. 1812: Publishes Dlsqulsitiones generales circa seriem infinitam, a rigorous treatment of infinite series, and introduces the hypergeometric function for the first time, for which he uses the notation F(et, fJ; Y; z): an essay on approximate integration. 1820-1830: Publishes over 70 papers, including Disquisitiones generales circa superficies curvas, in which he creates the intrinsic differential geometry of general curved surfaces,

416

14. COMPLEX ANALYSIS OF SOLOES

the forerunnerof Riemanniangeometryand the general theoryof relativity. From the 18308 on, Gauss was increasingly occupied with physics, and he enriched every branch of the subject he touched. In the theory of surface tension, he developed the fundamental idea of conservation of energy and solved the earliest problem in the calculus of variations. In op-

tics, he introducedthe conceptof the focal length of a systemof lenses.He virtuallycreated the science of geomagnetism, and in collaboration with his friend and colleague Wilhelm Weber he invented the electromagnetic telegraph. In 1839 Gauss published his fundamental paper on the general theory of inverse square forces, which established potential theory as a coherent branch of mathematics and in which he established the divergence theorem. Gauss had many opportunities to leave Gottingen, but he refused all offers and remained there for the rest of his life, living quietly and simply, traveling rarely, and working with immense energy on a wide variety of problems in mathematics and its applications. Apart from science andhis family-he married twice and had six children, two ofwhom emigrated to America-his main interests were history and world literature, international politics, and public finance. He owned alarge library of about 6000 volumes in many languages, including Greek, Latin, English, French, Russian, Danish, and of course German. His acuteness in handling his own financial affairs is shown by the fact that although he started with virtually nothing, he left an estate over a hundred times as great as his average annual income during the last half of his life. The foregoing list is the published portion of Gauss's total achievement; the unpublished and private part is almost equally impressive. His scientific diary, a little booklet of 19 pages, discovered in 1898, extends from 1796 to 1814 and consists of 146 very concise statements of the results of his investigations, which often occupied him for weeks or months. These ideas were so abundant and so frequent that he physically did not have time to publish them. Some of the ideas recorded in this diary: Cauchy Integral Formula: Gauss discovers it in 1811, 16 years before Cauchy. Non-Euclidean Geometry: After failing to prove Euclid's fifth postulate at the age of 15, Gauss came to the conclusion that the Euclidean form of geometry cannot be the only one possible. Elliptic Functions: Gauss had found many of the results of Abel and Jacobi (the two main contributors to the subject) before these men were born. The facts became known partly through Jacobi himself. His attention was caught by a cryptic passage in the Disquisitiones, whose meaning can only be understood if one knows something about elliptic functions. He visited Gauss on several occasions to verify his suspicions and tell him about his own most recent discoveries, and each time Gauss pulled 30-year-old manuscripts out of his desk and showed Jacobi what Jacobi had just shown him. After a week's visit with Gauss in 1840, Jacobi wrote to his brother, "Mathematics would be in a very different position if practical astronomy had not diverted this colossal genius from his glorious career." A possible explanation for not publishing such important ideas is suggested by his comments in a letter to Bolyai: "It is not knowledge but the act of learning, not possession but the act of getting there, which grants the greatest enjoyment. When I have clarified and exhausted a subject, then I turn away from it in order to go into darkness again." His was the temperament of an explorer who is reluctant to take the time to write an account of his last expedition when he could be starting another. As it was, Gauss wrote a great deal, but to have published every fundamental discovery he made in a form satisfactory to himself would have required several long lifetimes.

14.4 THE HYPERGEOMETRIC FUNCTION

417

A third relation can be obtained by making the snbstitntion w = (I-z)y-a- pu. Tltis leads to a hypergeometric equation for u with al = Y - a, fh = Y - fJ, and Yl = y. Furthermore, w is analytic at z = 0, and w(O) = I. We couclude that w = F(a, fJ; Y; z), We therefore have the ideutity F(a, fJ; Y; z) = (I - z)y-a- p F(y - a, Y - fJ; Y; z), (14.25) To obtain the canonical basis at z = I, we make the substitution t = 1- z, and notethattheresultisagaintheHGDE,withal = a,fJI = fJ,andYI = a+fJ-y+1. It follows from Equatiou (14.24) that

Y + I; 1- z), p (I - z)y-a- F(y - fJ, Y - a; Y - a - fJ + I; I - z)

W3(Z) sa F(a, fJ; a W4(Z) ==

+ fJ -

(14.26) form a canonical basis of solutious to the HGDE at z = I. A symmetry of the hypergeometric functiou that is easily obtained from the HGDEis (14.27)

F(a, fJ; Y; z) = F(fJ, a; Y; z).

The six functions F(a

± I, fJ; Y; z),

F(a, fJ ± I; Y; z),

F(a, fJ; Y

± I; z)

are called hypergeometric functions contiguous to F(a, fJ; Y; z), The discussiou above showed how to obtain the basis of solutions at z = I from the regular solutiou to the HDE z = 0, Fta, fJ; Y; z). We can show that the basis of solutious at z = 00 can also be obtained from the hypergeometric fuuctiou. Equatiou (14.16) suggests a function of the form v(z)

= z' F (ai, fJI; Yl;

D G) sa z'w

=}

w(z)

= z'v

G),

(14.28)

where r, ai, fJl' and Yl are to be determined. Since w(z) is a solution of the HGDE, v will satisfy the following DE (see Problem 14.15); z(l- z)v" + [I - a - fJ -2r - (2 - Y - 2r)z]v' 2

- [r - r

+ ry - ~(r + a)(r + fJ)] v =

O. (14.29)

This reduces to the HGDE if r = -a or r = -fJ. For r = -a, the parameters become al = a, fJl = I +a - Y, and YI = a - fJ + I. For r = -fJ, the parameters are al = fJ, fJI = I + fJ - Y, and Yl = fJ - a + I. Thus, VI(Z) = z-a F (a, I

+a -

Y; a - fJ + I;

V2(Z) = z-p F (fJ, I

+ fJ -

Y; fJ - a

+ I;

~) ,

D

(14.30)

418

14. COMPLEX ANALYSIS OF SOLDES

Kummer's solutions

form a canonical basis of solutions for the HGDE that are valid about z = 00. As the preceding discussion suggests, it is possible to obtain many relations among the hypergeomettic functions with different parameters and independent variables. In fact, the nineteenth-century mathematician Kummer showed that there are 24 different (but linearly dependent, of course) solutions to the HGDE. These are collectively known as Kummer's solutions, and six of them were derived above. Another important relation (shown in Problem 14.16) is that

z·-y (1 - z)Y-.-p F (Y - a, I - a; I - a

+ fJ; ~)

(14.31)

also solves the HGDE. Many of the functions that occur in mathematical physics are related to the hypergeomettic function. Even some of the common elementary functions can be expressed in terms of the hypergeometric function with appropriate parameters. For example, when fJ = y, we obtain . .

Fta,

Jacobi functions

fJ, fJ, z)

k _ _. _ ~ r(a + k) r(a)r(k + I) z - (I - z) .

f;;o

Similarly, F(~, ~; ~; z2) = sin- 1 zlz, and F(I, I; 2; -z) = In(1 + z)lz. However, the real power of the hypergeomettic function is that it encompasses almost all of the nonelementary functions encountered in physics. Let us look briefly at a few of these. Jacobi functions are solutions of the DE (I - x

2

d 2u

du

dx

dx

)-2 + [fJ - a - (a + fJ + 2)x]- +

).(). + a

+fJ +

I)u = 0 (14.32)

2z changes this DE into the HGDE with parameters al = )., + I, and Yl = I + a. The solutions of Equation (14.32), called the Jacobi functions of the first kind, are, with appropriate normalization,

Defining x

=I-

fJl = ). + a + fJ P

c. R) (z) = op

A

r(). + a + 1)

Z)

( IF -). ). + a + fJ + I' I + a' - - .

r(). + l)r(a + I ) '

"

2

When). = n, a nonnegative integer, the Jacobi function turns into a polynomial of degree n with the following expansion:

pCa, PJ(z) = n

r(n+a+l) r(n + I)r(n + a + fJ + I)

~r(n+a+fJ+k+I).(Z-I)k

f;;o

I'(o + k + I)

2

These are the Jacobi polynomials discussed in Chapter 7. In fact, the DE satisfied by pJ.,PJ (x) ofChapter7 is identical to Equation (14.32). Note that the transformation x = I - 2z translates the points z = 0 and z = 1 to the points x = I and x = -I, respectively. Thus the regular singular points of the Jacobi functions of the first kind are at ±1 and 00.

14.5 CONFLUENT HYPERGEOMETRIC FUNCTIONS

419

A second, linearly independent, solution of Equation (14.32) is obtained by using (14.31). These are called the Jacobi functions of the second kiud: lO,P) _ 2,,+o+Pr(A + ex + 1)r(A + fJ + I) Q, (z) - r(2A + ex + fJ + 2)(z - 1),,+o+l(z + I)P

. F (A+ex + I,A + 1; 2A +ex + fJ +2; _2_).

1-z

Gegenbauer functions

Gegeubauer functions, orultraspherical functions, are special cases ofJacobi functions for which ex = fJ = /-' - ~. They are defined by /L

_

c, (z) -

Legendre functions

(14.33)

1-Z)

r(A+2/-,) ( . 1. rCA + 1)r(2/-,) F -A, A+ 2/-" /-' + 2' -2- .

(14.34)

Note the change in the normalization constant. Linearly independent Gegenbauer functions "of the second kind" can be obtained from the Jacobi functions of the second kind by the substitution ex = fJ = /-' -~. Another special case of the Jacobi functions is obtained when ex = fJ = O. Those obtained from the Jacobi functions of the first kind are called Legendre functions of the first kind: F,(z)

_ (0,0) 1/2 = F, (z) = c, = F

(

Z)

. . 1-A, A + 1, 1, -2- .

(14.35)

Legendre fuuctions of the second kind are obtained from the Jacobi functions of the second kind in a similar way:

2'r

2(A

+ 1)

(

2 )

Q,(z) = r(2A + 2)(z _ 1)"+1 F A + 1, A + 1; 2A+ 2; 1 _ z .

Other functions derived from the Jacobi functions are obtained similarly (see Chapter7).

14.5 Confluent Hypergeometric Functions The transformation x = 1 - 2z translates the regular singular points of the HGDE by a finite amount. Consequently, the new functions still have two regular singular points, z = ±I, in the complex plane. In some physical cases of importance, only the origin, corresponding to r = 0 in spherical coordinates (typically the location of the source of a central force), is the singular point. If we want to obtain a differential equation consistent with such a case, we have to "push" the singular point z = 1 to infinity. This can be achieved by making the substitution t = rz in the HGDE and taking the limit r --> 00. The substitution yields 2w

(Y I - Y + ex + fJ) dw exfJ 0 d dt 2 + t + t- r + t(t _ r) w = .

at

(14.36)

420

14. COMPLEX ANALYSIS OF SOLOES

If we blindly take the limit r --+ 00 with a, fJ, and )I remaining finite, Equation (14.36) reduces to w+ ()I/t)th = 0, an elementary FODE in th. To obtain a nonelementary DE, we need to manipulate the parameters, to let some of them tend to infinity. We want )I to remain finite, because otherwise the coefficient of dw/dt will blow up. We therefore let fJ or a tend to infinity. The result will be the same either way because a and fJ appear symmetrically in the equation. It is customary to let fJ = r --+ 00. In that case, Equation (14.36) becomes

2w d dt 2 confluent hypergeomelric OE

+ (l'. _I) dw t

':W = t

O.

Multiplying by t and changing the independent variable back to Z yields zw"(z)

confluent hypergeometric function and series

_

dt

+ ()I -

z)w' (z) - OIW(Z) = O.

(14.37)

This is called the confluent hypergeometric DE (CHGDE). Since Z = 0 is still a regular singular point of the CHGDE, we can obtain expansions about that point. The characteristic exponents are 0 and 1-)1, as before. Thus, there is an analytic solution (corresponding to the characteristic exponent 0) to the CHGDE at the origin, which is called the confluent hypergeomelric function and denoted by
f'...

r(OI + k)

f;;Q r(k+ 1)r()I +k)z

k

.

(14.38)

This is called the conflnent hypergeometric series. An argument similar to the one given in the case of the hypergeometric function shows that

,

14.5.1. Box. The confluent hypergeometricfunction
A second solution of the CHGDE can be obtained, as for the HGDE. If 1 -)I is not an integer, then by taking the limit fJ --+ 00 of Equation (14.24), we obtain the second solution Zl-Y


-!

hydrogen-like atoms

V

2

1jf

+ ( 2E + -2Ze2) r - 1jf = O.

421

14.5 CONFLUENT HYPERGEOMETRIC FUNCTIONS

The radial part of this equation is given by Eqnation (12.14)with fer) Defining u = rR(r), we may write

= 2E + 2Ze21r.

e, - + (A + -a - -b) u = 0, dr 2

(14.39)

r2

r

where A = 2E, a = 2Ze2, and b = 1(1 + I). This equation can be further simplified by defining r == kz (k is anarbitrary constant to be determined later):

-d

2u

dz2

+ ( Ak2 + -ak - -b) u = O. Z

Choosing Ak2

z2

= -! and introducing a == al(2H) yields

d2u+(_!+~_~)u=0. 4 z z2 dz 2

Equations of this form canbe transformed intothe CHGDE by making thesubstitution u(z) = z"e- vz f(z).1t theu follows that 2f d +(2fJ. _2v)df dz 2 z dz Cboosing v2 =

I" + quantization ofthe energy ofthe hydrogen atom

+[_!+fJ.(fJ.- I )

! and fJ.(fJ. -

c:-

2v)

4 I)

_2fJ.v +~-~+v2]f=0. zzz2

= b reduces this equation to

r - 2fJ. zV

z2

a f = 0,

which is in the form of (14.37). On physical grounds, we expect u(z) --+ 0 as z --+ 00. 5 Therefore, v = Similarly, with fJ.(fJ. - I) = b = 1(1 + I), we obtain the two possibilities fJ. = -I and fJ. = I + I. Againonphysical grounds, we demand that u(O) be finite (thewavefunction must not blow up at r = 0). This implies'' that fJ. = I + I. We thus obtain

!.

fff

+ [2(1; I)

_ I]

t' _ I + ~ -

a f = O.

Multiplying by z gives zf" + [2(1 + I) - zlr" - (I + I - a)f = O. Comparing this with Equation (14.37) shows that f is proportional to cf>(I + I-a, 21 +2; z). Thus, the solution of (14.39) can be written as

u(z)

= Cz 1+ te- z/ 2cf> (1 + 1- a, 21 + 2; z).

An argument similar to that used in Problem 13.20 will reveal that the product e- z/ 2 cf> (I + I-a, 21+2; z) will be infinite unless the power series representing cf> terminates

(becomes a polynomial). It follows from Box 14.5.1 that this will take place if I

+ I-a =-N

(14.40)

5Thisis becausethevolumeintegral of 1\111 2 overall spacemustbe finite. Theradial part of thisintegral is simplytheintegral of r 2 R2(r) = u2(r). This latter integral will not be finite unless u(oo) = o. 6Recallthat JL is the exponent of z = r / k.

422

14. COMPLEX ANALYSIS OF SOLOES

for some integer N :::: O. In that case we obtain the Laguerre polynomials Lj = r(N+j+l) <1>(-N' N - I'(N + I)r(j + I) ,J

I')

+ ,z ,

where

j = 21 + I.

Condition (14.40) is the quantization rule for the energy levels of a hydrogen-like atom. Writing everything in tenus of the original parameters and defining n = N + I + 1 yields-after restoring all the m's and the !i's-the energy levels of a hydrogen-like atom: E = _ Z2 me4 = _Z2

21i2n 2

2

(me ) 2

a2~, n2

where a = eO 1(1ic) = 1/137 is the fine structure constant. The radial wave functions can now be written as Rn I(T) = Un,I(T) = CTle-bj(naO)<1> , r

where an

= 1i2 / (me2) = 0.529 x

(-n + 1 + 1,21 + 2; 2ZT) , naO

10-8 em is the Bohr radius.

III

Friedrich Wilhelm Bessel (1784-1846) showed no signs of unusual academic ability in school, although he did show a liking. for mathematics and physics. He left school intending to become a merchant's apprentice, a desire that soon materialized with a seven-year unpaid apprenticeship with a large mercantile finn in Bremen. The young Bessel proved so adept at accounting and calculation that he was granted a small salary, with raises, after only the first year. An interest in foreign trade led Bessel to study geography aod laoguages at uight, astonishingly learning to read and write English in only three months. He also studied navigation in order to qualify as a cargo officer aboard ship, but his innate curiosity soon compelled him to investigate astronomy at a more fundamentalleveI. Still serving his apprenticeship, Bessel learned to observe the positions of stars with sufficient accuracy to determine the longitude of Bremen, checking his results against professional astronomical journals. He then tackled the more formidable problem of determining the orbit of Halley's comet from published observations. After seeing the close agreement between Bessel's calculations and those of HaIley, the German astronomer Olbers encouraged Bessel to improve his already impressive work with more observations. The improved calculations, an achievement tantamount to a modem doctoral dissertation, were published with Olbers's recommendation. Bessel later received appointments with increasing authority at observatories near Bremen and in Konigsberg, the latter position being accompanied by a professorship. (The title of doctor, required for the professorship, was granted by the University of Gottingen on the recommendation of Gauss.) Bessel proved himself an excellent observational astronomer. His careful measurements coupled with his mathematical aptitude allowed him to produce accurate positions for a number of previously mapped stars, taking account of instrumental effects, atmospheric refraction, and the position and motion of the observation site. In 1820 he determined the position of the vernal equinox accurate to 0.01 second, in agreement with modern values.

14.5 CONFLUENT HYPERGEOMETRIC FUNCTIONS

423

His observation of the variation of the proper motion of the stars Sirius and Procyon led

him to posit the existence of nearby, large, low-luminosity stars called dark companions. Between 1821 and 1833he cataloguedthe positions of about 75,000 stars, publishing his measurements in detail. One of his most important contributions to astronomy was the determination of the distanceto a star using parallax. This method uses triangulation,or the determination of the apparent positions of a distant object viewed from two points a known distanceapart, in this case two diametrically opposedpoints of the Earth's orbit.The angle subtended by the baseline of Earth's orbit, viewed from the star's perspective, is known as the star's parallax. Before Bessel's measurement, stars were assumed to be so distant that their parallaxes were too small to measure, and it was further assumed that bright stars (thought to be nearer) would have the largest parallax. Bessel correctly reasoned that stars with large proper motions were more likely to be nearby ones and selected such a star, 61 Cygni, for his historic measurement. His measured parallax for that star differs by less than 8% from the currently accepted value. Given such an impressive record in astronomy, it seems only fitting that the famous functions that bear Bessel's name grew out ofhis investigations ofperturbations in planetary systems. He showed that such perturbations could be divided into two effects and treated separately: the obvious direct attraction due to the perturbing planet and an indirect effect caused by the sun's response to the perturber's force. The so-called Bessel functions then appear as coefficients in the series treatment of the indirect perturbation. Although special cases of Bessel functions were discovered by Bernoulli,Euler, and Lagrange, the systematic treatment by Bessel clearly established his preeminence, a fitting tribute to the creator of the most famous functions in mathematical physics.

14.5.1 Bessel Functions Bessel differential equalion

The Bessel differential equation is usually written as 2)

W 1/

I , + ( 1- -V + -w Z Z2

(14.41)

w= 0

As in the example above, the substitution w = inlo 2

d f

(2/k

+I

) df ~

-+---2~-+

di

z

which, if we set /k

[/k

2 -

v

2

~

zI'e-~' f(z)

- ~(2/kz+ I)

transforms (14.41)

2

+~+

IJ f =, 0

= v and ~ = i, reduces to

r + CV: I _ 2i) f' _ (2V; I); f = O. Making the further substitution 2; z = t, and multiplying out by t, we obtain

d 2f df t+(2v+I-t)--(v+!)f=O, 2 dt dt

424

14. COMPLEX ANALYSIS OF SOLOES

1

which is in the fonn of (14.37) with a = v + and y = 2v + 1. Thns, solutions of the Bessel equation, Equation (14.41), can be wtitten as constant multiples of zVe-iz(v + 2v + I; 2iz). With propernonnalization, we define the Bessel function of the first kind of order v as

1,

Bessel function ofthe firstkind

Jv(z) =

1 f(v

+ 1)

(Z)v. 1 e-"(v + 2' 2v 2

+ 1; 2iz).

(14.42)

Using Equation (14.38) and the expansion for e- iz, we can show that

Z ) V 00 (_I)k (Z)2k Jv(z)= ( 2" ~k!f(V+k+l) 2"

(14.43)

.

The second linearly independent solution can be obtained as usual and is proportional to

ZI-(2v+1)

=C

Bessel function ofthe second kind, or Neumann function

G) G)

v e-iz(v

+ 1- (2v + I) + I, 2 ., (2v + I); 2iz)

-v e-iz(_v +

1, -2v + 1; 2iz) = C1- v(z) ,

provided that 1- y = 1- (2v + 1) = -2v is not an integer. When v is an integer, 1-n(z) = (_I)n In(z) (see Problem 14.25). Thus, when v is anoninteger, the most general solution is of the form AJv(z) + B1- v(z). How do we find a second linearly independent solution when v is an integer n? We first define

Yv(z) =

J» (z) cos V1f - 1- v(z) . , sm uzr

(14.44)

called the Bessel function of the second kind, or the Nenmann function. For noninteger v this is simply a linear combinatiou of the two linearly independent solutious. For integer v the function is indeterminate. Therefore, we use I'Hopital's rule and define

Yn(z)

. Yv(z) = -1 == v-+n Ion

1{

v . [aJ Ion - (-1) na1-V) -- . 8v

v--+-n

8v

Equation (14.43) yields

(!:) _ (!:)V f( _1)k W(v + k + 1)

aJv = Jv(z) In av 2

2

k=O

k!f(v

+ k + 1)

(!:)2k, 2

where W(z) = (dldz) In I'(z), Sintilarly,

(!:) + (!:)-v f

a1- v = -1- v(z) In av 2

2

W(-v

+ k + 1)

k~O k!f(-v +k+

(!:)2k. 1) 2

425

14.5 CONFLUENT HYPERGEOMETRIC FUNCTIONS

Substituting these expressions in the definition of Yn (z) and using L n (z) = (-I)nJn(z), we obtain

2 Yn(z)=;Jn(z)ln

(Z) I (z)n~ k qr(n+k+l) (Z)2k 2: -; 2: 6(-1) k!I'(n+k+l) 2:

_~(~)-n -I n~ -I 2

T(

Bessel function ofthe third kind orHankel function

()

qr(k-n+l) 6( ) k!I'(k - n + I) k

er 2

(14.45)

The natural log term is indicative of the solution suggested by Theorem 14.2.6. Since Yv(z) is linearly independent of Jv(z) for any v, integer or noninteger, it is convenientto consider {Jv(z), Yv(z)} as a basis of solutions for the Bessel equation. Another basis of solutions is defined as (14.46) which are called Bessel functions of the third kind, or Hankel functions. Replacing z by iz in the Bessel equation yields

whose basis of solutions consists of multiples of Jv(iz) and Lv(iz). Thus, the modified Bessel fnnctions of the first kind are defined as I

-

v (z) = e

.

-mvj2],

. u (zz) -

(Z)v I (Z)2k 2: t; k!I'(v + k + I) 2: . 00

Similarly, the modified Bessel fnnctions of the second kind are defined as T(

Kv(z) = 2'

SIll

v:rr

[I-v(z) - Iv(z)].

When v is anintegern, In = L n, and K n is indeterminate. Thus, we define Kn(z) as limv-->n Kv(z). This gives

Kn(z) = (_I)n lim [aL v _ aIv]. 2 v--+n Bv Bv which has the power-series representation

426

14. COMPLEX ANALYSIS OF SOLOES

recurrence relation for solutions of the BesselOE

We can obtain a recorrence relation for solntions of the Bessel eqnation as follows. If Zv(z) is a solntion of order v, then (see Problem 14.28) and

v d v Zv-l = C2C -[z Zv(z)].

dz

If the constants are chosen in snch a way that Zv, Z-v, Zv+l, and Zv-l satisfy their appropriate series expansions, then Cl = -I and C2 = 1. Carrying ont the differentiation in the equations for Zv+l and Zv-l, we obtain

v

d Z;

z

z

v

Zv-l = -Zv

Zv+l = -Zv - - d '

z

sz; + --. dz

(14.47)

Adding these two eqnations yields the recursion relation Zv-l(Z)

+ Zv+l(Z) =

2v

(14.48)

-Zv(z),

z

where Zv(z) can be any of the three kinds of Bessel functions.

14.6 Problems 14.1. Show that the solution of w'

+ w/ Z2

= 0 has an essential singularity at

z =0. 14.2. Derive the recursion relation of Equation (14.7) and express it in terms of the indicial polynomial, as in Equation (14.9). 14.3. Find the characteristic exponent associated with the solution of ui" + p(z)w ' q(z)w = 0 at an ordinary point [a point at which p(z) and q(z) have no poles]. How many solutions can you find?

+

14.4. The Laplace equation in electrostatics when separated in spherical coordinates yields a DE in the radial coordinate given by

~ (x 2 dY ) dx

dx

_ n(n

+ I)y = 0

for n 2: O.

Starting with an infinite series of the form (14.6), show that the two independent solutions of this ODE are of the form x n and x-n - 1 .

14.5. Find the indicial polynomial, characteristic exponents, and recursion relation at both of the regular singular points of the Legendre equation, wI! -

2z a _-w' + -w = O. I - Z2

l - z2

What is ai, the coefficient of the Laurent expansion, for the point z =

+ I?

14.6 PROBLEMS

427

14.6. Show that the substitution z = lit transforms Equation (14.13) into Equation (14.14). 14.7. Obtain the indicial polynomial of Equation (14.14) for expansion about t = O. 14.8. Show that Riemann DE represents the most general second order Fuchsian DE. 14.9. Derive the indicial equation for the Riemann DE. 14.10. Show that the transformation v(z) = ZA(Z - I)"w(z) changes the pairs of characteristic exponents 0.1, A2), (iLl, iL2), and (VI, V2) for the Riemann DE to (AI + A, A2 + A), (J.L1 + u; iL2 + iL), and (VI - A - u, V2 - A - iL).

14.11. Go through the steps leading to Equations (14.24), (14.25), and (14.26). 14.12. Show that the elliptic function of the first kind, defined as K(z) =

de

" /2

/0o

JI- Z2 sin 2 e

,

can be expressed as (n 12)F(~, ~; I; Z2).

14.13. By differentiating the hypergeometric series, show that dn r(a + n)r(fJ + n)r(y) dz n F(a, fJ; y; z) = r(a)r(fJ)r(y + n) F(a

+ n, fJ + n; y + n; z).

14.14. Use direct substitution in the hypergeometric series tu show that F(-a, fJ; fJ; -z) = (I

+ z)",

I F(I, I; 2; -z) = -In(1

z

1 1. 3. 2 1'_1 F(2' 2' 2'Z) = -sm z,

z

+ z).

14.15. Show that the substitution v(z) = ZT w(l/z) [see Equation (14.28)] transforms the HGDE into Equation (14.29). 14.16. Consider the function v(z) '" zT(I - z)' F(al, fJI; Yl; liz) and assume that it is a solution of HGDE. Find a relation among r, s, aI, fJt, and Yl such that v(z) is written in terms of three parameters rather than five. In particular, show that one possibility is v(z) = za- y (I - z)y-a-P F(y - a, I - a; I

Find all such possibilities.

+ fJ -

a; liz).

428

14. COMPLEX ANALYSIS OF SOLOES

14.17. Show that the Jacobi functions are related to the hypergeometric fooctions. 14.18. Derive the expression for the Jacobi function of the second kind as given in Equation (14.33). 14.19. Show that z =

00

is not a regular singular point of the CHGDE.

14.20. Derive the confluent hypergeometric series from hypergeometric series. 14.21. Show that the Weber-Hermite equation, u" +(v+ ~ - !Z2)u = ocan be transformed into the CHGDE. Hint: Make the substitution u(z) = exp( Z2)v(Z).

-!

14.22. The linear combination .

_

ljI(a, y, z) =

I'(l - y) . I) (a, y, z) I'(a - y +

+

I'(y - I) l-y _ I'(a) Z (a y j, a

+ 1,2 -

.

y, z)

is also a solntion of the CHGDE. Show that the Hermite polynomials can be written as

2 Z) " nlz H" (.,fi = 2 1jI(-2' 2; 2)' 14.23. Verify that the error fooction erf'(z) = J~ e-t'dt satisfies the relation erf'(z) = z(~, ~; _Z2). 14.24. Derive the series expansion of the Bessel function of the first kind from that of the confluent hypergeometric series and the expansion of the exponential. Check your answer by obtaining the same result by snbstituting the power series directly in the Bessel DE. 14.25. Show that 1-,,(z) = (-I)"J,,(z). Hint: Let v J v (z) and use I'Cm) = 00 for a nonpositive integer m,

= -n in the expansion of

14.26. In a potential-free region, the radial part of the Schrodinger equation reduces to 2

d R dr 2 spherical Bessel functions

+ ~ dR + [I.. _ :!...] R = 2 r dr

O.

r

Write the solutions of this DE in terms of Bessel functions. Hint: Substitute R = uj.p. These solutions are called spherical Bessel functions.

14.27. Theorem 14.2.6 states that ooder certain conditions, linearly independent solutions of SOLDE at regular singolar points exist even though the difference between the characteristic exponents is an integer. An example is the case of Bessel fooctions of half-odd-integer orders. Evaluate the Wronskian of the two linearly independent solutions, Jv and J- v , of the Bessel equation and show that it vanishes only ifv is an integer. This shows, in particular, that J,,+1/2 and 1-,,-1/2 are linearly independent. Hint: Consider the value of the Wronskian at z = 0, and use the formula I'(v)I'(l - v) = x] sin vrr.

14.6 PROBLEMS

429

14.28. Show that z±V(djdz)[z'FvZv(z)] is a solution of the Bessel equation of order v ± I if Z, is a solution of order v. 14.29. Use the recursion relation of Equation (14.47) to prove that

I d)m u lz Zv(z)] = zv- m Zv-m(Z), ( ~ dz I d

(~dJ

m

[z-VZv(z)] = (-I)mz-V-mZv+m(z).

14.30. Using the series expansion of the Bessel function, write JI/2(Z) and Ll/2(Z) in terms of elementary functions. Hint: First show that r(k + ~) =

v'IT (2k + 1)!j(k!22k+l).

14.31. From the results of the previous two problems, derive the relations

L n - I/2(Z) =

[fr

l 2 /

I n+I/2(Z) = [fzn+I/2

(~:y ("o;Z) , (_~:y ("i~Z) .

14.32. Obtain the following integral identities: (a)

(b) (e)

f f f

zV+l Jv(z)dz = zV+l Jv+l(Z). z-v+IJv(z)dz = -z-v+lJv_l(Z). ZIL+l Jv(z)dz

=

ZIL+

1

Jv+l(Z)

+ (I-' -

f

- (1-'2 - v 2)

ZIL-

v)ZIL Jv(z) 1

Jv(z) dz;

and evaluate (d)

f

Z3 Jo(z)

dz:

Hint: For (c) write ZIL+ 1 = ZIL-Vz v+1 and use integration by parts. 14.33. Use Theorem 14.2.6 and the fact that In(z) is entire to show that for integer n, a second solution to the Bessel equation exists and can be written as Yn (z) = I n (z)[/n (z) + en ln z], where In(z) is analytic about z = O. 14.34. (a) Show that the Wronskian W(Jv, Z; z) of J v and any other solution Z of the Bessel equation, satisfies the eqnation d dz[zW(Jv, Z; z)] =

o.

430

14. COMPLEX ANALYSIS OF SOLOES

(b) For some constant A, show that ~ [~] _ W(z) _ _ A_ dz J v - J;(z) - zJ;(z)'

(c) Show that the general second solution of the Bessel equation can be written as

14.35. Spherical Bessel functions are defined by

Let il(Z) denote a spherical Bessel function "of some kind." By direct differentiation and substitution in the Bessel equation, show that

(c) Combine the results of parts (a) and (b) to derive the recursion relations il-t(Z)

+ fi+l(z)

2/ + 1 = --fj(z),

/fi-I (z) - (/

z

+ l)fi+1 (z) =

(2/

d f;

+ 1) dz .

2 14.36. Show that W(Jv, Yv ; z) = 2/(:n:z), W(HJ!), HS ); z) = 4/(i:n:z). Hint: Use Problem 14.34.

14.37. Verify the following relations: (a) Yn+I/2(Z)

(b) Lv(z)

= (_1)n+1 L

n-I/2(Z),

= sin v:n: Jv(z) + cos v:n:Yv(z) =

(c) Y-n(z) = (-l)nyn(z)

L n-I/2(Z) = (-1)" I n+I/2(Z). J v (z) - cos v:n: Lv (z)

.

sm vn

.

in the limit v --> n in part (b).

14.38. Use the recurrence relation for the Bessel function to show that JI (z) = -Jij(z). 14.39. Let u = J v ().z) and v = J v (jLz). Multiply the Bessel DE for u by v [z and that of v by ufz, Subtract the two equations to obtain

0.2 _

jL2)zuv =

~

dz

[z (u dv _ v dU)] . dz dz

(a) Write the above equation in terms of J v (}"z) and J v (jLz) and integrate both sides with respectto Z.

14.6 PROBLEMS

431

(b) Now divide both sides by A2 - ,,2 and take the limit as" --> A. You will need to use L'Hopital's rule. (c) Substitute for J;'(AZ) from the Bessel DE and simplify to get

f

z[Jv(Az)fdz =

Z; {[j;(AZ)]2 + ( 1- A~:2) [Jv(Az)f} .

(d) Finally, let X = xvn/a, where r.; is the nth root of J v, and use Equation (14.47) to arrive at

{a

2

(X vn

)

a

22

io zJv ---;;z dz = 2Jv+I (x vn). 14.40. The generating function for Bessel functions of integer order is exp[~z(t­ I/t)]. To see this, rewtite the generating function as e"/2e- z/21, expand both factors, and wtite the product as powers of t", Now show that the coefficient of t" is simply In(z). Finally, use Ln(z) = (-I)nJn(z) to derive the formula exp[~z(t - I/t)] = L~-oo JIl(z)t n. 14.41. Make the substitutions z = fJtY and w = tau to transform the Bessel DE 2 into t 2 d u + (2", + l)t du + (fJ2y 2 t2Y + ",2 _ v 2y 2)u = O. Now show that dt 2

dt

Airy's differential equation ii - tu = 0 has solutions of the form hf3(~it3/2) and LI/3(~it3/2). d2 w

14.42. Show that the general solution of dt 2 t[AJv(e l/')

o is

+

w =

+ BYv(e l/')].

14.43. Transform dw/dz + w 2 + zm = 0 by making the substitution w = (d/dz) In v. Now make the forther substitutions

v= u..;z

and

t =

_2_ z !+O/2)m

m+2

to show that the new DE can be transformed into a Bessel equation of order I/(m + 2). 14.44. Starting with the relation

exp[~x(t - l/t)] exp[~Y(t - l/t)] = exp[~(x

+ y)(t -

l/t)]

and the fact that the exponential function is the generating function for I n (z), prove the "addition theorem" for Bessel functions: 00

JIl(x

+ y)

=

L k=-oo

h(X)Jn-k(Y)·

432

14. COMPLEX ANALYSIS OF SOLDES

Additional Reading 1. Birkhoff, G. and Rota, G.-C. Ordinary Differential Equations, 3rded., Wiley, 1978. The first two sections of this chapter closely follow their presentation. 2. Dennery, P. and Krzywicki, A. Mathematics for Physicists, Harper and Row, 1967. 3. Watson, G. A Treatise on the Theory of Bessel Functions, 2nd ed., Cambridge University Press, 1952. As the name suggests, the definitive text and reference on Bessel functions.

15

_

Integral Transforms and Differential Equations

The discussion in Chapter 14 introduced a general method of solving differential equations by power series-i-also called the Frobenius method-which gives a solution that converges within a circle of convergence. In general, this circle of convergence may be small; however, the function represented by thepowerseries can be analytically continued using methods presented in Chapter II. This chapter, which is a bridge between differential equations and operators on Hilbert spaces (to be developed in the next part), introduces another method of solving DEs, which uses integral transforms and incorporates the analytic continuation automatically. The integral transform of a function v is another function u given by u(z) =

kernel of integral transforms examples of integral transforms

fc

K(z, t)v(t)dt,

(15.1)

where C is a convenient contour, and K(z, t), called the kernel of the integral transform, is an appropriate function of two complex variables. 15.0.1. Example. Let us consider some examples of integral transforms. (a) The Fouriertransform is familiar from the discussionof Chapter 8. The kernel is

K(x, y) = e ixy. (b) The Laplace transform is used frequently in electrical engineering. Its kernel is

K(x, y) = e- xy.

(c) The Euler transform has the kemel K(x, y) = (x _ y)v.

434

15. INTEGRAL TRANSFORMS AND DIFFERENTIAL EQUATIONS

(d) The Mellin transform has the kernel

st», y) = G(x Y). where G is an arbitrary function. Most of the time K(x, y) is taken to be simply x Y.

(e) The Hankel transform has the kernel K(x, y) = yJn(xy), where In is the nth-order Bessel function. (f) A transform that is useful in connection with the Bessel equation has the kernel

K(x,y)

Strategy forsolving DEs using integral transforms

=

Gf

e y-x'/4y.

IIIi

The idea behind using integral transform is to write the solution u(z) of a DE in z in terms of an integral such as Equation (15.1) and choose v and the kernel in such a way as to render the DE more manageable. Let Lz be a differential operator (DO) in the variable z. We want to determine u(z) such that Ldu] = 0, or equivalently, such that Ie Lz[K (z, t)]v(t) dt = O. Snppose that we can find M" a DO in the variable t, such that LdK(z, t)l = M,[K(z, t)]. Then the DE becomes Ie(M,[K(z, t)]Jv(t) dt = O. If C has a and b as initial and final points (a and b may be equal), then the Lagrange identity [see Equation (13.24)] yields

0= Lz[u] =

t

K(z, t)M;[V(t)]dt

+ Q[K, v]I~,

where Q[K, v] is the "surface term." If vet) and the contour C (or a and b) are chosen in such a way that Q[K, v]l~ = 0

and

M;[V(t)] = 0,

(15.2)

the problem is solved. The trick is to find an M, such that Equation (15.2) is easier to solve than the original equation, Lz [u] = O. This in torn demands a clever choice of the kernel, K (z, t). This chapter discusses how to solve some common differential equations of mathematical physics using the general idea presented above.

15.1 Integral Representation of the Hypergeometric Function Recall that for the hypergeometric function, the differential operator is

d2 Lz = z(1 - z) dz 2

+ [y

- (a

d

+ fJ + I)zl dz

- afJ

For such operators-whose coefficient functions are polynomials-the proper choice for K(z, t) is the Euler kernel, (z - t)'. Applying Lz to this kernel and

15.1 INTEGRAL REPRESENTATION OF THE HYPERGEOMETRIC FUNCTION

435

rearrmgmgrenns,weO~IDn

Lz[K(z, t)] = /Z2[-s(s - 1) - s(a +,B + 1) - a,B] +st(a +,B

+ 1) + 2a,Bt] -

+ z[s(s- 1) + sy

yst - a,BP}(z - t)'-2. (15.3)

Note that except for a multiplicative constant, Kiz; t) is synunetric in z and t. This suggests that the general form of M, may be chosen to be the same as that of Lz except for the interchange of z and t. If we can manipulate the parameters in such a way that M, becomes simple, then we have a chance of solving the problem. For instance, if M, has the form of Lz with the constant term absent, then the hypergeometric DE effectively reduces to a FODE (in dv/ dt). Let us exploit this possibility. The general form of the M, that we are interested in is

d2 M, = P2(t) dt 2

d

+ PI (t) dt'

i.e., with no Po term. By applying M, to K(z, t) = (z - t)' and setting the result equal to the RHS of Equation (15.3), we obtain

s(s - 1)p2 - PISZ + pist = Z2[-s(s - 1) - s(a +,B + z[s(s - 1) + sy - yst - a,Bt2,

+ 1) - a,B] + st(a + ,B + 1) + 2a,Bt]

for which the coefficients of equal powers of z on both sides must be equal:

+ 1) - a,B = 0 =} s = -a 1) + sy + stt« +,B + 1) + 2a,Bt,

- s(s - 1) - s(a +,B - PIS = s(s s(s - 1)p2

+ pvst =

or

s = -,B,

-yst - a,Bt2.

If we choose s = -a (s = -,B leads to an equivalent representation), the coefficient functions of M, will be completely determined. In fact, the second equation gives PI (t), and the third determines P2(t). We finally obtain PI(t) =a+l-y +t(,B -a -1), and 2

2 d M, = (t - t )-2 dt

+ [a + 1 -

y

+ t(,B -

d a - 1)]-, dt

(15.4)

which, according to Equation (13.20), yields the following DE for the adjoint:

Mi[v]

d2

= -dt2 [(t -

d t 2)v] - -{[a - y dt

+ 1 + t(,B -

a - l)]v}

= O. (15.5)

436

15. INTEGRAL TRANSFORMS AND DIFFERENTIAL EQUATIONS

The solution to this equation is v(t) = Cta-Y(t- I)Y-~-l (see Problem 15.5). We also need the surface tenm, Q[K, v], in the Lagrangeidentity (see Problem 15.6 for details): Q[K, v](t) = Cat a-y+1(t - l)Y-~(z - t)-a-l. Finally, we need a specification of the contour. For different contours we will get different solutions. The contour chosen must, of course, have the property that Q[K, v] vanishes as a result of the integration. There are two possibilities: Either the contour is closed [a = b in (15.2)] or a i' b but Q[K, v] takes on the same value at a and at b. Let us consider the second of these possibilities. Clearly, Q[K, v](t) vanishes att = 1 ifRe(y) > Re(,8). Also, as t ---> 00,

Q[K, v](l) --->

(_l)-a-1Cata-y+ltY-~t-a-l

= (_l)-a-1Cat-~,

which vanishes if Re(,8) > O. We thus take a = 1 and b = 00, and assume that Re(y) > Re(,8) > O. It then follows that

=

u(z)

l

b

K(z, t)v(t)dt

=c

1

00

(t - z)-ata- y (t -

l)y-~-ldt. (15.6)

The constaot C' Can be determined to be r(y)/[r(,8)r(y - ,8)] (see Problem 15.7). Therefore,

u(z) sa F(a,,8; y; z) = Euler formula for fhe hypergeomelric function

r(,8)~i~) _

1

00

,8)

y (t - z)-ata- (t

-l)y-~-ldt.

It is customary to change the variable of integration from t to 1/ t. The resulting expression is called the Euler formula for the hypergeometric function:

F(a ,8' y' z) = ,

"

r(y) t(l r(,8)r(y - ,8) 0

1

tz)-at~-l(l- t)y-~-ldt. (15.7)

Note that the tenm (1 - tz)-a in the integral has two branch points in the zplane, one at z = l/t and the other at z = 00. Therefore, we cut the z-plane from Zl = l/t, a point on the positive real axis, to Z2 = 00. Since 0 :'0 t :'0 1, Zl is somewhere in the interval [1, (0). To ensure that the cutis applicable for all values of t, we take Zl = 1 and cut the plane along the positive real axis. It follows that Equation (15.7) is well behaved as long as

o < arg(l -

z) < 21f.

(15.8)

We could choose a different contour, which, in general, would lead to a different solution. The following example illustrates one such choice.

15.1.1.Example. Firstnotethat Q[K, v] vanishes att = Oandt = 1 as long as Re(y) > Re(li) and Re(a) > Re(y) -I. Hence, we can choose the contour to start att = 0 and end

437

15.2 INTEGRAL REPRESENTATION OF THE CONFLUENT HYPERGEOMETRIC FUNCTION

at t

=

1. We thenhave

(15.9)

Tosee therelation between w(z) andthe hypergeometric function, expand (1- t/z)-a in the integral to get w(z) = e"z-u

f

r(a

+ n)

(~)n [I t u +n-y (1 _

n~O f(a)r(n + 1) z

tly-P-1dt.

Jo

(15.10)

Now evaluatethe integralby changing t to l/t and osing Eqoations (11.19) and (11.17). This changes the integral to

[00 t-u-n-l+P(t _ l)y-P- 1dt

Jl

= r(a

+ n + 1 - y)r(y r(a + n + 1 - fJ)

fJ)

Snbstitnting this in Equation (15.10), we obtain

c"

c"

1- y) (1)"

-a ~ f(a +n)f(a +n + F'(e + n + 1 - fJ)r(n

w(z) = r(a) r(y - fJ)z

= r(a) f(y - fJ)z-a

f;;n

+ 1) Z

r(a)r(a + 1- y) r(a + 1- fJ) F(a, a - y

+ 1; a -

fJ + 1; l/z),

where we have usedthe hypergeometric series of Chapter 14.Choosing

e" =

f(a + 1 - fJ) r(y - fJ)r(a + 1 - y)

yields w(z) = z-a F(a, a - y + 1; a - fJ + 1; l/z), which is one of the solutions of the II!I hypergeometric DE [Equation (14.30)].

15.2 Integral Representation of the Confluent Hypergeometric Function Having obtained the integral representation of the hypergeometric fnnction, we can readily get the integral representation of the conflnent hypergeometric function by taking the proper limit. It was shown in Chapter 14 that (a, y; z) = limp-+oo F(a, fJ; y; zlfJ). This snggests taking the limit of Equation (15.7). The presence of the ganuna functions with fJ as their arguments complicates things, but on the other hand, the symmetry of the hypergeometric function can be utilized

438

15. INTEGRALTRANSFORMS AND DIFFERENTIAL EQUATIONS

to our advantage. Thus,we may write <1>(a, y; z) =

lim F (a,

~ ....oo

= lim

fJ;

y; .:.) =

fJ

r

r(y)

~ ....co r(a)r(y - a) Jo

= <1>(a, y; z) =

(fJ, a; y;

lim F

~

.... oo

(1-

.:.)

fJ

IZ)-~ 1.-1(1 _ fJ

t)y-.-Idl

t

r(y) eztl·-I(I _ t)y-.-Idl r(a)r(y - a) Jo (15.11)

because the limit of the firsttenn in the integral is simply e tz . Note that the condition Re(y) > Re(a) > 0 must still hold here. Integral transforms are particularly useful in determining the asymptotic behavior of functions. We shall use them in deriving asymptotic formulas for Bessel functions later on, and Problem 15.10 derives the asymptotic formula for the confluent hypergeometric function.

15.3 Integral Representation of Bessel Functions Choosing the kernel, the contour, and the function v(l) that lead to an integral representation of a function is an art, and the nineteenth century produced many masters of it. A particularly popular theme in such endeavors was the Bessel

equation and Bessel functions. This sectionconsiders theintegral representations of Bessel functions. The most effective kernel for the Bessel DE is K(z,

I) Gr (I - ~:). =

exp

Z

When the Bessel DO Lz sa

L K(z, t) z

Z)

d + -I d V -z + ( 1- -z dz

z dz

z

.

acts on K(z, I), it yields

t z'/4 ZZ)(Z)V (d V+I) = ( -v+1 -I - + I + -41 z -2 e - t = -dt - -I - K(z ' I).

Thus, Mt = dfdt - (v t dv Mt [V(I)] = - - dt

+ 1)/1, and Equation (13.20) gives v +I - - v = 0, I

whose solution, including the arbitrary constant ofintegrationk, is v(l) = ta:":", When we substitute this solution and the kernel in the surface term of the Lagrange identity, Equation (13.24), we obtain Q[K, V](I) = PIK(z, l)v(l) = k

G)

v

l- v-I et - z' / (4t ) .

15.3 INTEGRAL REPRESENTATION OF BESSEL FUNCTIONS

439

Imt

Ret The contourC in the r-planeused in evaluating Jv(z).

Figure 15.1

A contour in the I -plane that ensures the vanishing of Q [K, v]for all values of v starts at I = -00, comes to the origin, orbits it on an arbitrary circle, and finally goes back to t = -00 (see Figure 15.1). Such a contour is possible because of the factor et in the expression for Q[K, v]. We thus can write Jv(z) = k

inlegral representation of Bessel function

G) fc v

t-V-Iet-z2/(4t)dt

(15.12)

Note that the integrand has a cut along the negative real axis due to the factor I-v-I. If v is an integer, the cut shrinks to a pole att = O. The constant k must be determined in such a way that the above expression for Jv(z) agrees with the series representation obtained in Chapter 14. It can be shown (see Problem 15.11) that k = 1/(2,,;). Thus, we have Jv(z) =

~ (~)V 2",

2

r

lc

t-v-let-z2/(41)dt

It is more convenient to take the factor (zI2) v into the integral, introduce a new integration variable u = 2t [z, and rewrite the preceding equation as Jv(z) =

r

~

2", Jc

u- v- 1e(z/2)(u-lfu)du.

(15.13)

This result is valid as long as Re(zu) < 0 when u --> -00 on the negative real axis; that is, Re(z) must be positive for Equation (15.13) to work. An interesting result can be obtained from Equation (15.13) when v is an integer. In that case the only singularity will be at the origin, so the contour can be taken to be a circle about the origin. This yields J (z) = _1_ n 2ni

Bessel generating function

r

Jc

u-n- Ie(z/2)(u-lfu)du

'

which is the nth coefficient of the Laurent series expansion of exp[(zI2)(u -l/u)] about the origin. We thus have this important resnlt:

L

DO

e(z/2)(t-l/l) =

n=-oo

In(z)t n.

(15.14)

440

15. INTEGRAL TRANSFORMS AND DIFFERENTIAL EQUATIONS

The function exp[ (zIZ) (t - 1It) 1is therefore appropriately called the generating fnnction for Bessel functions of integer order (see also Problem 14.40). Equation (15.14) can be useful in deriving relations for such Bessel functions as the following example shows. 15.3.1. Example. Let us rewrite the LHS of (15.14) as eZI/2e-z/21, expand the exponentials, andcollect terms to obtain

f

e(z/2)(t-ljt) = ezt/2e-z/2t =

~ (~)m 2

m=O m.

=

ff

f

~ (_~)n

n=O n.

2t

(-~)~ (~)m+n t m-n. 2

m=On=O m.n.

If we let m - n = k, change the m summation to k, andnote that k goes from we get

e(z/2)(I-I/I) =

f f

(_I)n

k~-oon~O (n +k)!n!

k~oo 00

[(Z)k

2:

-00

to 00,

(~)2n+k rk

E 00

r(n

2 (_I)n

(Z)2n] k

+ k + l)r(n + 1) 2:

1 .

Comparing thisequation withEquation (15.14)yieldsthefamiliar expansion fortheBessel function: Z) k

h(z) = ( 2:

E

(_I)n

00

r(n

(Z)1.

+ k + l)r(n + 1) 2:

.

Wecanalsoobtain arecurrence relationfor I n (z), DifferentiatingbothsidesofEquation (15.14) with respect to t yields

f

~ (1 + k)e(Z/2)(I-I/I) =

nJn(Z)ln- 1

(15.15)

n=-oo

Using Equation (15.14) on the LHS gives n L G+ --;) In(z)t = n=-oo 2t 00

~

00

00

L In(z)t n + ~ n=-oo L I n(z)t n-2 n=-oo 00

L

=~

n=-oo

In_l(Z)t''-I+~

00

L n=-oo

I n+I(Z)I,,-I, (15.16)

where we substituted n - 1for n in thefirst sumandn + 1 for n inthesecond. Equating the coefficientsof equal powers of 1 on the LHS and the RHS of Equations (15.15) and (15.16), we get

z

nJn(z)= 2:[Jn-I(Z)

+ J,,+I(Z)],

which was obtained hy a different method in Chapter 14 [see Eq. (14.48)].

III

15.3 INTEGRAL REPRESENTATION OF BESSEL FUNCTIONS

441

Imw

Rew

The contour C' in the w-plane used in evaluating JvCz).

Figure 15.2

We can start with Equation (15.13) and obtain other integral representations of Bessel functions by making appropriate substitutions. For instance, we can let u = e" and assume that the circle of the contour C has unit radius. The contour C' in the w-plane is determined as follows. Write u = rei() and w sa x + iy, sol re iO = eXe iy yielding r = eX and eiO = e iy. Along the first part of C, = -rr and r goes from 00 to 1. Thus, along the corresponding part of C', y = -rr and x goes from 00 to O. On the circular part of C, r = 1 and e goes from -rr to -l-z , Thus, along the corresponding part of C', x = 0 and y goes from -rr to +rr. Finally, on the last part of C', y = rr and x goes from 0 to 00. Therefore, the contour C' in the w-plane is as shown in Figure 15.2. Substituting u = eW in Equation (15.13) yields

e

Jv(z) = ---.!:..., ( ezsmhw-vwdw,

lei

21rl

Re(z) > 0,

(15.17)

which can be transformed into (see Problem 15.12)

In" n

Jv(z) = -1

sin vrr cos(ve -zsine)de _ -

0

integral representation of Bessel functions of integer order

1r

In 0

oo

. 'dt. e- vt- z sinh (15.18)

For the special case of integer v, we obtain I n (z) =

~

(" cos(ne - z sin e) de. rr Jo

In particular, Jo(z) = -1 rr

In"

cos(z sine) de.

0

We can use the integral representation for Jv (z) to find the integral representation for Bessel functions of other kinds. For instance, to obtain the integral 1Do not confuse x and y with the realandimaginary parts of z,

442

15. INTEGRAL TRANSFORMS AND DIFFERENTIAL EQUATIONS

Imw

in

1------1~----

en Rew Figure15.3 The contour c" in the w-planeused in evaluating

HS 1) (z).

representation for the Neumann function Yv(z), we use Equation (14.44):

1 Yv(z) = (cot V1f)Jv(z) - -.--Lv(z) sm V1r 1C cot _ V1l' cos VJr = _ cos(ve-zsine)de---

i

tc

0

- .1 tt

sm Vll'

i"

e- Vf -

z Sinh

"dt

0

7i

cos(ve +zsine)de _ _1

0

i"" i"" .

1C

evt-zsinhtdt

0

with Re(z) > O. Substitute tc - e for e in the third integral on the RHS. Then insert the resulting integrals plus Equation (15.18) in HJl) (z) = Jv(z) + iYv(z) to obtain

Re(z) > O. These integrals can easily be shown to result from integrating along the contour e" of Figure 15.3. Thus, we have

Re(z) > O. By changing i to

-t, we can show that Re(z) > 0,

where

e

lll

is the mirror image of

e" about the real axis.

15.4 ASYMPTOTIC BEHAVIOR OF BESSEL FUNCTIONS

443

15.4 Asymptotic Behavior of Bessel Functions As mentioned before, integral representations are particularly useful for determining the asymptotic behavior offunctions. For Bessel functions we can consider two kinds of limits. Assumiug that both v and z = x are real, we can consider v --+ 00 or x --+ 00. First, let us consider the behavior of Jv (x) oflarge order. The appropriate method for calculating the asymptotic form is the method of steepest descent discussed in Chapter II for which v takes the place of the large parameter a. We use Equation (15.17) because its integrand is simpler than that ofEquation (15.13). The form of the integrand in Equation (15.17) may want to suggest f(w) = -w and g (w) = e" sinh w. However, this choice does not allow setting f' (w) equal to zero. To proceed, therefore, we write the exponent as v

(~ sinh

w- w), and

conveniently introduce x]» es 1/ cosh wo, with wo a real number, which we take to be positive. Substituting this in the equation above, we can read off sinh w

f(w) =

cos

h

wo

g(w) = 1.

-w,

°

The saddle point is obtained from df/dw = or cosh w = cosh woo Thus, w = ±wo + 2in:n:, for n = 0, 1, 2 .... Since the contour C ' lies in the right halfplane, we choose Wo as the saddle point. The second derivative fN (wo) is simply tanh wo, which is real, making Ih = 0, and el = n /2 or 3:n:/2. The convention of Chapter II suggests taking el = :n: /2 (see Figure 15.4). The rest is a matter of substitution. We are interested in the approximation to w up to the third order in t: w - wo = bIt + b2t 2 + b3t3. Using Equations (11.31), (11.37), and (11.38), we can easily find the three coefficients:

v"i

bl =

eWI = i

1f"(woJll/2

=

b 2

f"'(wo) e4i81

=

31f"(wo)12

v"i

,Jtanh wo' 2 cosh Wo 3 sinh 2 wiJ'

5[f"'(wo)f f(4l(w O)} v"ie 3i91 b3 = { 3[f"(wO)]2 - f"(wo) 121f"(wo)!3/2 -

-

.

-I

v"i

12(tanh wO)3/2

(53 coth2 Wo -

1)

.

If we substitute the above in Equation (11.36), we obtain the following asymptotic formula valid for v --+ 00:

Jv(x)'"

eX(sinh WO-Wij cosh WO)

(2:n:x sinh wo)I/2

[

1+

I

8x sinh Wo

where v is related to WQ via v = x cosh woo

(

5 2) ] I--coth wo + ... 3 '

444

15. INTEGRAL TRANSFORMS AND DIFFERENTIAL EQUATIONS

v ire---------

u

-ix - - Figure 15.4

~-=--=:-::=- -

-

Thecontour Co in the w-plane used in evaluating Jv(z) for large values of

v. Let us now consider the asymptotic behavior for large x. It is convenient to consider the Hankel functions HSt)(x) and HS2)(x). The contours e" and ell! involve both the positive and the negative real axis; therefore, it is convenient, assuming that x > v, to write v = x cos fJ so that

The saddle points are given by the solutions to cosh W = cos fJ, which are ±ifJ. Choosing WQ = +ifJ, we note that the contour along which Im(sinh W

- W

cos fJ) = Im(sinh WQ

-

WQ

WQ

=

cos fJ)

is given by coshu = [sinfJ + (v - fJ)cosfJ]/sinv. This contour is shown in Figure 15.5. The rest of the procedure is exactly the same as for Jv(x) described above. In fact, to obtain the expansion for HS1) (x), we simply replace WQ by ifJ. The result is

H(l)(x) '" (.

v

2

.

mx smfJ

)

V2

ei(xsin~-v~)

I 5 [I + g.x. smfJ . (I + -cot2 fJ) + ... J. 3

When x is much larger than v, fJ will be close to ](/2, and we have

HS1l(x) '" (2e i (X-

'In

which, with

11x

vrr / 2- rr/ 4)

(1 + ~) , g.x

-+ 0, is what we obtained in Example 11.5.2.

15.5 PROBLEMS

445

v

u Thecontour in the w-plane used in evaluating .H51) (z) in thelimitof large

Figure 15.5 values of x.

The other saddle point, at -if!, gives the other Hankel function, with the asymptotic limit H(2)(X) '" (2e- i(X-vrr/2-rr/4)

V;;

v

(1 _~) 8IX

We can now use the expressions for the asymptotic forms of the two Hankel functions to write the asymptotic forms of Jv(x) and Yv(x) for large x: Jv(x) =

&[H~l)(X) + H~2\x)]

'" (2 [cos (x

V;-;

;i [H~l)(x) "'J :x

Yv(x) =

-

- v~2 - ~) +~ sin (x - v~ - ~) + ... ] ' 4 8x 2 4 H~2)(x)]

[Sin(x-vi

-~) - 8~ cos (x-vi -~)+"l

15.5 Problems 15.1. Use the change of variables k = Int and ix = w -Cl (wherek and x are the common variables used in Fourier transform equations) to show that the Fourier transform changes into a Mellin transform, G(t) = -I . 21l'1

l

ioo u

+ F(w)t-Wdw,

-ioo+a

15.2. The Laplace transform L[f] of a function f(l) is defined as L[J](s) sa

10

00

e- st f(t) dt.

446

15. INTEGRAL TRANSFORMS AND DIFFERENTIAL EQUATIONS

Show that the Laplace transform of (a)

f(t) = 1

is

(b)

f(t) = cosh tot

is

(c)

f(t) = sinhwt

is

(d)

f(t) =coswt

is

(e)

f(t) = sinwt

is

(f)

f(t) = eM for t > 0,

is

(g)

f(t) = t"

is

I

where s > O.

s

s

s2 -

w2 '

w s2 -

(J)2'

s s2 +w2· w s2 + w 2· 1

,

s-w f(n + 1) , sn+l

where s 2 >

(li2.

where s > w. where s > 0, n > -1.

15.3. Evaluate the integral f(t) =

00

10o

sin tat --dw

w

by finding the Laplace transform and changing the order of integration. Express the result for both t > 0 and t < 0 in terms of the theta function. (You will need some results from Problem 15.2.)

15.4. Show that the Laplace transform of the derivative of a function is given by L[F'](s) = sL[F](s) - F(O). Similarly, show that for the second derivative the transform is L[FI/](s) = s2L[F](s) - sF(O) - F'(O). Use these results to solve the differential equation. ul/(t) the boundary conditions u(O) = a, u'(O) = O.

+ w 2u (t ) =

0 snbject to

15.5. Solve the DE of Equation (15.5). 15.6. Calculate the surface term for the hypergeometric DE. 15.7. Determine the constant C' in Equation (15.6), the solution to the hypergeometric DE. Hint: Expand (t - z)-a inside the integral, use Equations (11.19) and (11.17), and compare the ensuing series with the hypergeometric series of Chapter 14. 15.8.: Derive the Euler formula [Equation (15.7)]. 15.9. Show that F(a,

fJ;

f(y)f(y - a y; 1) = f(y _ a)r(y -

fJ) fJ)'

(15.19)

447

15.5 PROBLEMS

Hint: Use Equation (11.19). Equation (15.19) was obtained by Gauss using only hypergeometric series.

15.10. We determine the asymptotic behavior of (a, y; z) for z --> 00 in this problem. Break up the integral in Equation (15.11) into two parts, one from to -00 and the other from -00 to 1. Substitute -t/z for t in the first integral, and 1 - t / z for t in the second. Assuming that z --> 00 along the positive real axis, show that the second integral will dominate, and that

o

r(y)

(a, y; z) --> f(a)z·-YeZ

as

z ----*

00.

15.11. In this problem, we determine the constant k of Equation (15.12). (a) Write the contour integral of Equation (15.12) for each of the three pieces of the contour. Note that arg(t) = -n as t comes from -00 and arg(t) = n as t goes to -00. Obtain a real integral from 0 to 00. (b) Use the relation f(z)f(1 - z) = tt] sin n z, obtained in Chapter 11, to show that

n

I'(v-z) =

r(z + 1) sin xz

(c) Expand the function exp(z2/4t) in the integral of part (a), and show that the contour integral reduces to

., ~ (z)2n f(-n - v) sm VJr L." .

-2l

n=O 2

f(n

+ 1)

(d) Use the resnlt of part (c) in part (b), and compare the resnlt with the series expansion of Jv(z) in Chapter 14 to arrive finally at k = 1/(2ni).

15.12. By integrating along Ct, C2, C3, and C4 of Figure 15.2, derive Equation (15.18). 15.13. By substituting t = exp(ill) in Equation (15.14), show that eiZ';ne = Jo(z)

00

00

n=l

n=O

+ 2 I: hn(z) cos(2nll) + 2i I: hn+J (z) sin[(2n + 1)11].

In parricular, show that Jo(z) = - 1

2n

121< e,
15.14. Derive the integral representations of HJl)(x) and HJ2) (x) given in Section 15.3.

448

15. INTEGRAL TRANSFORMS AND DIFFERENTIAL EQUATIONS

Additional Reading 1. Dennery, P. and Krzywicki, A Mathematicsfor Physicists, Harper and Row, 1967.

2. Mathews, J. and Walker, R. Mathematical Methods of Physics, 2nd ed., Benjamin, 1970.

Part V

Operators on Hilbert Spaces

_

16,

_

An Introduction to Operator Theory

The first two parts of the book dealt almost exclusively with algebraic techniques. The third and fourth part were, devoted to analytic methods. In this introductory chapter, we shall try to unite these two branches of mathematics to gaiu insight into the nature of some of the important equations in physics and their solutions. Let us start with a familiar problem.

16.1 From Abstract to Integral and Differential Operators Let's say we wantto solve an abstract vector-operator equation A lu) = Iv) in an Ndimensional vector space V. To this end, we select a basis B = {lUi) )~1' write the equation in matrix form, and solve the resulting system of N linear equations. This produces the components of the solution lu) in B. If components in another basis B' are desired, they can be obtained using the similarity transformation connecting the two bases (see Chapter 3). There is a standard formal procedure for obtaining the matrix equation. It is convenient to choose an orthonormal basis B = IIei) } ~1 for V and refer all components to this basis. The procedure involves contracting both sides of the equation with (eil and inserting 1 = r.7=1Iej) (ejl between A and lu):

r.7=1 (eil A lej) (ejl u) = (eil v) for i = 1,2, ... , N, or N

LAijuj j=l

= Vi

fori=1,2, ... ,N,

(16.1)

452

16. AN INTRODUCTION TO OPERATOR THEORY

where Ai} sa (eilAlej) , Uj ss (ejlu), and Vi ss (eilv). Equation (16.1) is a system of N linear equations in N unknowns Iu j J7=1' which can be solved to obtain the solution(s) of the original eqnation in B. A convenient basis is that in which A is represented by a diagonal matrix diag(AI, A2, '" ,AN). Then the operator equation takes the simplcform x.a, = Vi, and the solution becomes immediate. Let us now apply the procedure just described to infinite-dimensional vector spaces, in particnlar, for the case of a continuous index. We want to find the solutions of K Iu) = If). Following the procedure used above, we obtain (xl K

(lab IY) w(y) (YI dY) lu} = lab (xl K Iy) w(y) (YI u) dy = (xl f) ,

where we have used the results obtained in Chapter 6. Writing this in functional notation, we have

lab K(x, y)w(y)u(y) dy = f(x), Integral operators and kernels

local operators

(16.2)

which is the continuous analogue of Equation (16.1). Here (a, b) is the interval on which the functions are defined. We note that the indices have tumed into continuous arguments, and the snm has turned into an integral. The operator K that leads to an equation such as (16.2) is called an integral operator (10), and the "matrix element" K (x, y) is said to be its kernel. The discussion of the discrete case mentioned the possibility of the operator A being diagonal in the given basis B. Let us do the same with (16.2); that is, noting that x and y are indices for K, let us assume that K(x, y) = 0 for x oft y. Such operators are called local operators. For local operators, the contribution to the integral comes ouly at the point where x = y (hence, their name). If K (x, y) is finite at this point, and the functions w (y) and u (y) are well behaved there, the LHS of (16.2) will vanish, and we will get inconsistencies. To avoid this, we need to have K (x,y ) =

o {00

if x oft y, . If x = y.

Thus, K(x, y) has the behavior of a delta function. Letting K (x, y) sa L(x)8(x y)/w(x) and substituting in Equation (16.2) yields L(x)u(x) = f(x). Inthediscrete case, A.i wasmerely anindexed number; itscontinuous analogue, L(x), may represent merely a function. However, the fact that x is a continuous variable (index) gives rise to other possibilities for L(x) that do not exist for the discrete case. For instance, L (x) could be a differential operator. The derivative, although defined by a limiting process involving neighboring points, is a local operator. Thus, we can speak of the derivative of a function at a point. For the

16.2 BOUNDED OPERATORS IN HILBERT SPACES

453

discrete case, u; can only "hop" from i to i + I and then back to i. Snch a difference (as opposed to differential) process is notlocal; it involves not only i but also i + 1. The "point" i does not have an (infinitesimally close) neighbor. This essential difference between discrete and continuous operators makes the latter far richer in possibilities for applications. In particular, if L(x) is considered a differential operator, the equation L(x)u(x) = f(x) leads directly to the fruitful area of differential equation theory.

16.2 Bounded Operators in Hilbert Spaces The concept ofan operator on a Hilbert space is extremely subtle. Even the elementary characteristics of operators, such as the operation of hermitian conjugation, cannot generally be defined on the whole Hilbert space. In finite-dimensional vector spaces there is a one-to-one correspondence between operators and matrices. So, in some sense, the study of operators reduces to a stody of matrices, which are collections of real or complex numbers. Although we have already noted an analogy between matrices and kernels, a whole new realm of questions arises when Aij is replaced by K(x, y)-questions about the continuity of K(x, y) in both its arguments, about the limit of K(x, y) as x and/or y approach the "end points" of the interval on which K is defined, about the boundedness and "compactness" of K, and so on. Such subtleties are not unexpected. After all, when we tried to generalize concepts of finite-dimensional vector spaces to infinite dimensions in Chapter 5, we encountered difficulties. There we

were concerned aboutvectors only; the generalization of operators is even more complicated.

right-shift operator

16.2.1. Example. Recall that
T, = T, (f>j la)

j=l

lej ))

=

f>j Iej+I).

j=l

In other words, Tr transforms ("'10 "'2 •... ) to (0, "'I. "'2, ... ). It is straightforward to show

that Tr is indeed a linear operator.

..

The first step in our stody of vector spaces of infinite dimensions was getting a handle on the convergence of infinite sums. This entailed defining a norm for vectors and a distance between them. In addition, we noted that the set of linear transformations L (V, W) was a vector space in its own right. Since operators are "vectors" in this space, the study of operators requires constructing a norm in L (V, W) when V and W are infinite-dimensional.

454

16. AN INTROOUCTION TO OPERATOR THEORY

16.2.2. Definition. Let 1fl and 1fz be two Hilbert spaces with norms II . 111 and II· liz. ForanyT EL(1fl,1fz),thenumber max { II

TXllzl IIxlll

operator norm

bounded operator

lx}

'i- o}

(if it exists) is called 1 the operator norm ofT and is denoted by IITII. A linear transformation whose norm is finite is called a bounded linear transformation. A bounded linear transformation from a Hilbert space to itself is called a bounded operator. The collection ofall bounded linear transformations, which is a subset of L(1f" 1fz), will be denoted by 13(1fl, 1fz), and if1fl = 1fz es 1f, it will be denoted by 13(1f). Note that II . 111 and II . liz are the norms induced by the inner product of 1fl and 1fz. Also note that by dividing by IIx II t we eliminate the possibility of dilating the norm of IITII by choosing a "long" vector. By restricting the length of [x), one can eliminate the necessity for dividing by the length. In fact, the norm can equivalently be defined as IITII

= max IIlTxllzl

IIxlll

= I} = max IIlTxllzl

IIxliJ

="

I}.

(16.3)

It is straightforward to show that the three definitions are equivalent and they indeed define a norm. 16.2.3. Proposition. An operator T is bounded if and only if it maps vectors of finite norm to vectors offinite norm.

Proof Clearly, ifT is bounded, then IITxll has finite norm. Conversely, if IITxllz is finitefor all [x) (ofunitlength),maxIIlTxllzlllxllt = I} is also finite, andTis bounded. D An innnediate consequence of the definition is

IITxllz =" IITII IIxllJ 'of Ix) E 1f,.

(16.4)

Ifwechoose Ix}-Iy} instead of [x), it will follow from (16.4) that as [x) approaches Iy}, T [x) approaches T Iy}. This is the property that characterizes continuous functions: bounded operators arecontinuous

16.2.4. Proposition. The bounded operatorT tion from 1f, to 1fz.

E

13(1fj, 1fz) is a continuousfunc-

Another consequence of the definition is that 1The precise definition uses "supremum" instead of"maxiroum." Rather than spending a lot of effort explaining the difference between the two concepts, we use the less precise, but more intuitively familiar, concept of "maximum."

455

16.2 BOUNDEO OPERATORS IN HILBERT SPACES

16.2.5. Box. :B(:J-Cj, :J-Cz) is a vector subspace of!'(:J-Cj, :J-Cz), andfar:J-C, = :J-Cz = :J-C, we have 1 E 13(:J-C) and 11111 = I.

16.2.6. Example. Wehaveseen thatin aninnerproduct space,one canassociate a linear operator (linear functional) to everyvector. Thus,associated with thevector Ix} in a Hilbert

spaceXts thelinearoperatorIx : :J£ -e- C defined by Ix (Iy» '" (x] y). Wewantto compare theoperator normoffx with the norm of Ix}. Firstnote thatby usingthe Schwarz inequality, we get

IIlxll =max[

Ilxl~/)IIIY) "'0) =max[ I(~~~)IIIY) "'0) s

114

On the otherhand, from IIxli Z = Ix (Ix», we obtain

IXI~~I) ~ max [ Ilxl~I»IIIY) '" 0)

IIxll =

=lllxll·

Thesetwo inequalities implythat IIlx II = IIxli. derivative operator is unbounded

III

16.2.7. Example. The derivative operator D

dJdx is not a bounded operator on

the Hilbert space- £}(a, b) of sqnare-integrable fnoctions. With a function like f(x) = Jx - a, one gets b-a "" Ilfll = while df[dx = 1/(2.,fx - a) gives 1I0fllZ = 11011 = 00.

norm ofa product is less than the product of norms.

.Ji '

! J% dx/(x -

a) =

00.

We conclude that III

16.2.8. Example. Since £,O{) is an algebra as well as a vectorspace, one may be interested in the relation betweenthe product of operators andtheir norms. More specifically, one may want to knowbow IISTII is relatedto IISII and IITII. Iu this example we sbowthat

IISTII

~

IISII II TII.

(16.5)

Todo so, we use the definition of operator nonn for the product ST:

IISTII = max [

II~:~II

I Ix) '" 0)

=max[ IISTxll IITxll I Ix) "'O",Tlx)) IITxll IIxll Txlll SCT IX)1I1 < max [ II Tlx) ",o)max[ II lx) "'0). - . IITxll ~---,I,-Ix,-II ~_~ =IITII 2Herethe two Hilbertspaces coincide, so thatthe derivative operator acts on a single Hilbertspace.

456

16. AN INTROOUCTION TO OPERATOR THEORY

Now note that the first term on the RHS does not scan all the vectors for maximality: It scans only the vectors in the image of T. If we include all vectors, we may obtain a larger number. Therefore,

max { 11

5(Tlx))111 T1x);"0) < max { 115.< 111 Ix) ;"0) = 11511 IITxll IIxll '

andthe desiredinequalityis established. A usefulconsequenceofthisresultis IITn II ::; IITll n, which we shall use frequently.

IlIIII

We can put Equation (16.5) to immediate good use.

16.2.9. Proposition. Let 1f be a Hilbert space and T

E

13(1f). If [T II < I, then

1 - T is invertible and (1 - T)-t = L~o Tn.

Proof. First note that the series converges, because

Il

n=O f

Tn n=O n~O

Tn l1 s f

II

II :::: f

n IITli = I_I T II II

and the sum has a finite norm. Furthermore, (1 -

n f T n = (1 n=O

= lim k-+oo

T)

(lim tTn) k->;oon=O

= lim (1 - T) k--+oo

(trn -trn+t) n=O

n=O

trn

n=Q

= lim (1 - Tk+l) = 1, k-HXi

because 0 < Iim,....co IITk+ 1 11 ::; limr....oo IITIlk+l = 0 for IITII < I, and the vanishing of the norm implies the vanishing of the operator itself. One can similarly show that (L~o Tn )(l - T) = 1. D A corollary of this proposition is that operators that are "close enough" to an invertible operator are invertible (see Problem 16.1). Another corollary, whose proof is left as a straightforward exercise, is the following:

16.2.10. Corollary. Let T E 13(1f) and A a complex number such that IITII < IAI. Then T - Al is an invertible operator, and lOOT n (T - Al)-t = - A n~O A

L (-)

Adjoints play an important role in the stody of operators. We recall that the adjoint ofT is defined as (yl T Ix)* = (r] rt Iy) or (Txl y) = (x ITt y).Inthefinitedimensional case, we could calculate the matrix representation of the adjoint in a particular basis using this definition and generalize to all bases by similarity transformations. That is why we never raised the question of the existence of the adjoint of an operator. In the infinite-dimensional case, one must prove such an existence. We state the following theorem without proof:

16.3 SPECTRA OF LINEAR OPERATORS

457

16.2.11.Theorem. Let T E 1I(:Ji). Then the adjoint ofT, defined by (Txl y) = (x ITt y), Tand Tt have equal

exists. Furthermore, IITII = IITt II.

norms

Another useful theorem that we shall use later is the following. 16.2.12. Theorem. Let N(T) and range ofT E 1I(:Ji). We have

~(T)

denote the null space (kernel) and the

Proof Ix} is in N(Tt) iff rt [x) = 0 iff (y ITtx) = 0 for all Iy} E :Ji. This holds if and only if (Ty Ix) = 0 for all Iy) E :Ji. This is equivalent to the statement that

Ix} is in ~(T)-L. This chain of argumentproves that N(Tt) = ~(T)-L. The second part of the theorem follows from the fact that (Tt)t = T. D

16.3 Spectra of Lineal" Operators One of the most important results of the theory of finite-dimensional vector spacesis the spectraldecomposition theoremdevelopedin Chapter4.The infinitedimensional analogue of that theorem is far more encompassing and difficult to prove. It is beyond the scope of this book to develop all the machineryneeded for a thoroughdiscussionof the infinite-dimensional spectraltheory.Instead,weshall present the central results, and occasionallyintroduce the reader to the peripheral arguments when they seem to have their own merits. regular point ofan operator resolvent setand spectrum ofan operator

every eigenvalue of an operator on a

vector space offinite dimension isin its spectrum and vice versa

16.3.1. Definition. Let T E .t:,(:Ji). A complex number A is called a regular point ofT if the operator T - A1 is bounded and invertible.3 The set ofall regularpoints ofT is called the resolvent set ofT, and is denoted by p(T). The complement of p(T) in the complex plane is called the spectrum ofT and is denoted by a(T). Corollary16.2.10implies"that ifT is bounded,then p (T)is not empty, andthat the spectrum of a bounded linear operator on a Hilbert space is a boundedset.In fact, an immediate consequenceof the corollary is that A::; IITII for all A E aCT). It is instructive to contrast the finite-dimensional case against the implications of the above definition. Recall that because of the dimension theorem, a linear operator on a finite-dimensional vector space V is invertible if and only if it is either onto or one-to-one.Now, A E a(T) if and only ifT - A1 is not invertible. For finite dimensions, this implies thatS ker(T - A1) 01 O. Thus, in finite dimensions,

3IfT is bounded, then T- ).1is automatically bounded. 40ne can simply choosea x whoseabsolute value is greater than IITII. SNote how critical finite-dimensionality is for this implication. In infinite dimensions, an operator can be one-to-one (thus having a zerokernel) without beingonto.

458

16. AN INTRODUCTION TO OPERATOR THEORY

A E ,,(T) if and only if there is a vector la) in V such that (T - A1) la) = O. This is the combined definition of eigenvalue and eigenvector, and is the definition we will have to use to define eigenvalues in infinite dimensions. It follows that in the finite-dimensional case, ,,(T) coincides with the set of all eigenvalues of T. This is not true for infinite dimensions, as the following example shows.

notall points of aCT) areeigenvalues

16.3.2. Example. Consider the right-shift operator acting on Coo. It is easy to see that IITrall = lIall for all ]«). This yields IITrll = I, so that any A that belongs to a(T,) must be suchthat 1)..1 :::: 1. Wenow show that the converse is also true, i.e., that if III :5 1, then )" E o (T,.). It is sufficient to show that if 0 < 1),,1 ::: I, then Tr - A1 is not invertible. To establish this,we shallshowthat Tr - A1 is not onto. Suppose that T; -),,1 is onto. Then there must be a vector la) such that (Tr-),,1) la) = let} where leI} is the first standard basisvector ofCco. Equating components onbothsides yields therecursion relations al = -1/)", anda j-l = )"aj for all j :::: 2. One canreadily solvethisrecursion relation to obtain a j = -1/).) forall j. Thisis a contradiction, because 00

00

1

I: lajl2 = J=l I: IAI2j J=l will not converge ifO < IAI ~ 1, i.e., la} f/. Coo, and therefore Tr - A1 is not onto. we conclude thet c rr-) = {A E Cj O < IAI::: I). 1fwe could generalize the result of thefinite-dimensional case to Coo, we wouldconclude that all complexnumbers whose magnitude is atmost I are eigenvalues of Tr- Quiteto oursurprise, thefollowingargument showsthat Tr has no eigenvalues at all! Suppose that A is an eigenvalue of Tr. Let la} be any eigenvector for A. Since T r preserves the tength of a vector, we have (al a) = (Tral Tra) = (Aal Aa) = IAI2 (al a). It follows that IAI = I. Nowwrite la) = {aj )1=1 and let am be the first nonzeroterm of this sequence. Then 0 = (Tral em) = (Aul em) = Aum • The first equality comes about because T; la} has its first nonzero term in the (m + l)st position. Since A =1= 0, we must II!I have am = 0, whichcontradicts thechoice of thisnumber.

16.4 Compact Sets This section deals with some technical concepts, and as such will be rather formal. The central concept of this section is compactoess. Although we shall be using compactness sparingly in the sequel, the notion has sufficient application in higher analysis and algebra that it warrants an introductory exposure. Let us start with the familiar case of the real line, and the intuitive notion of "compactness." Clearly, we do not want to call the entire real line "compact," because intuitively, it is not. The next candidate seems to be a "finite" interval. So, first consider the open interval (a, b). Can we call it compact? Intuition says "yes," but the following argument shows that it would not be appropriate to call the open interval compact. Consider the map e : IR ---> (a, b) givenby etr) = b"2a tanh t+bi a • The reader may check that this map is continuous and bijective. Thus, we can continuously

16.4 COMPACT SETS

459

map all of R in a one-to-one manner onto (a, b). This makes (a, b) "look" very much" like lR. How can we modify the interval to make it compact? We do not want to alter its finiteness. So, the obvious thing to do is to add the end points. Thus, the interval [a, b] seems to be a good candidate; and indeed it is. The next step is to generalize the notion of a closed, finite interval and eventually come up with a definition that can be applied to all spaces. First we need some terminology. open ball

16.4.1. Definition. An open ball B; (x) of radius r and center Ix) in a normed vector space V is the set ofall vectors in V whose distance from Ix} is strictly less than r:

Br(x) open round neighborhood

== lly) EVilly -xII

< r].

We call Br(x) an open round neighborhood of [x), This is a generalization of open interval because

(a, b) =

{y

E

lR

Ily _a: b I <

b; a} .

16.4.2. Example. A prototype of finite-dimensional nonned spacesis JRn. An openball of radius r centered at x is

Br(x) = {y

E

JR I (Yt - Xt)2

+ (Y2 -

X2)2

+ ... + (Yn -

xn)2 <

r 2 }.

Thus, all points inside a circle form an open ball in the xy-plane, and all interior points of

a solid sphere form an open baIl in space.

bounded subset

II

16.4.3. Definition. A bounded subset of a normed vector space is a subset that can be enclosed in an open ball offinite radius. For example, any region drawn on a piece of paper is a bounded subset of lR2 , and any "visible" part of our environment is a bounded subset of lR3 because we can always find a big enough circle or sphere to enclose these subsets.

open subset boundary point closed subsetand closure

16.4.4. Definition. A subset (')ofa normed vector space V is called open if each of its points (vectors) has an open round neighborhood lying entirely in ('). A boundary point of(') is a point (vector) in V all ofwhose open round neighborhoods contain points inside and outside ('). A closed subset e of V is a subset that contains all of its boundary points. The closure ofa subset S is the union of S and all of its boundary points, and is denoted by S. For example, the boundary of a region drawn on paper consists of all its boundary points. A curve drawn on paper has nothing but boundary points. Every point is also its own boundary. A boundary is always a closed set. In particular, a point is a closed set. In general, an open set cannot contain any boundary points. A frequently used property of a closed set e is that a convergent sequence of points of e converges to a point in e.

6In mathematical jargon one says that (a, b) and R are homeomorphic.

460

16. AN INTROOUCTION TO OPERATOR THEORY

dense subset

16.4.5. Definition. A subset W 01 a normed vector space V is dense in V if the closure olW is the entire space V. Equivalently, W is dense if each vector in W is infinitesimally close to at least one vector in V. In other words, given any !u) E V and any E > 0, there is a !w) E W such that lIu - wll < E, i.e., any vector in V can be approximated, with arbitrary accuracy, by a vector in W.

rational numbers are dense inthe real numbers

A paradigm of dense spaces is the set of rational numbers io the normed vector space of real numbers. It is a well-known fact that any real number can be approximated by a rational number with arbitrary accuracy: The decimal (or bioary) representation of real numbers is precisely such an approximation. An iotuitive way ofimagioiog denseness is that the (necessarily) infinite subsetis equal to almost all of the set, and its members are scattered "densely" everywhere io the set. The embeddiog of the rational numbers in the set of real numbers, and how they densely populate that set, is a good mental picture of all dense subsets. A useful property iovolviog the concept of closure and openness has to do with contiouous maps betweennormedvectorspaces. Let I: 1t:1 --> 1t:2be a continuous map. Let (92 be an open setin1t:2. Let 1-1((92) denote the ioverse image of (92,i.e., all poiots of1t:1 that are mapped to (92. Let !XI) be a vector io I-I ((92), !X2) = I(!XI)), and let B.(X2) be a ball contained entirely io (92. Then I-I (B. (X2)) contaios !XI) and lies entirely io 1-1((92). Because of the continuity of I, one can now construct an open ball centered at IXI) lyiog entirely io I-I (B. (X2)), and by inclusion, in I-I ((92). This shows that every poiot of I-I ((92) has a round open neighborhood lyiog entirely io 1- 1((92) . Thus, 1-1((92) is an open subset. One can similarly show the correspondiog property for closed subsets. We can summarize this io the followiog: 16.4.6. Proposition. Let I : 1t:1 --> 1t:2 be continuous.Then the inverse image 01 an open (closed) subset 011t:2 is an open (closed) subset ol1t:J.

p(T) is open, and u(T) is closed and bounded inC.

Consider the resolvent set of a bounded operator T. We claim that this set is open io C. To see this, note that if )" E P (T), then T - )"1 is iovertible. On the other hand, Problem 16.1 shows that operators close to an invertible operator are iovertible. Thus, if we choose a sufficiently small positive number E and consider all complex numbers I-' withio a distance E from )", then all operators of the form T - 1-'1 are iovertible, i.e., I-' E P (T). Therefore, any)" E peT) has an open round neighborhood io the complex plane all poiots of which are io the resolvent. This shows that the resolvent set is open. In particular, it cannot contain any boundary poiots. However, peT) and O'(T) have to be separated by a common boundary? Sioce p (T) cannot contaio any boundary poiot, 0' (T) must carry the entire boundary. This shows that 0' (T) is a closed subset of Co Recalliog that 0' (T) is also bounded, we have the following result.

7The spectrum of a bounded operator need not occupy any "area" in the complex plane. It may consist of isolated points or line segments, etc., in which case the spectrum will constitute the entire boundary.

16.4 COMPACT SETS

461

16.4.7. Proposition. For any T E 1l(9i) the set peT) is an open subset ofC and is a closed, bounded subset of Co

IT (T)

Let us go back to the notion of compactness. It tums out that the feature of the closed interval [a, b] most appropriate for generalization is the behavior of infinite sequences of numbers lying in the interval. More specifically, let la,1;';;1 be a sequence of infinitely many real numbers all lying in the interval [a, b]. It is intuitively clear that since there is not enough room for these points to stay away from each other, they will have to crowd around a number of points in the interval. For example, the sequence

in the interval [-1, +1] crowds around the two points -t and +t.ln fact, the points with even n accumnlate around + and those with odd n crowd around It tums out that all closed intervals of ~ have this property, namely, all sequences crowd around some points. To see that open intervals do not share this property consider the open interval (0, 1). The sequence 12n~tl:l = Il,~, ... I clearly crowds only around zero, which is not a point of the interval. But we already know that open intervals are not compact.

t

compact subset

t.

16.4.8. Definition. (Bolzano-Weierstrass property) A subset X ofa normed vector space is called compact if every (infinite) sequence in X has a convergent subsequence. The reason for the introduction of a subsequence in the definition is that a sequence may have many points to which it converges. But no matter how many of these points there may exist, one can always obtain a convergent subsequence by choosing from among the points in the sequence. For instance, in the example

above,one can choose the subsequenceconsisting ofelementsfor which n is even.

t.

This subsequence converges to the single point + An important theorem in real analysis characterizes all compact sets in ~n:8 16.4.9. Theorem. (BWHB theorem) A subset of~n is compact if and only closed and bounded.

aCT) is compact

if it is

We showed earlier that the spectrum of a bounded lioear operator is closed and bounded. Identifying C with ~2, the BWHB theorem implies that

BBWHB stands for Balzano, Weierstrass, Heine, and Borel. Balzano and Weierstrass proved that any closed and bounded subset of R has the Balzano-Weierstrass property. Heine and Borel abstracted the notion of compactness in terms of open sets, and showed that a closed bounded subset of R is compact. The BWHB theorem as applied to R is usually called the HeineBorel theorem (although some authors call it the Balzano-Weierstrass theorem). Since the Balzano-Weierstrass property and compactness are equivalent, we have decided to choose BWHB as the name of our theorem.

462

16. AN INTROOUCTION TO OPERATOR THEORY

16.4.10. Box. The spectrum ofa bounded linear operator is a compact sub-

set of C. An immediate consequence of the BWHB Theorem is that every bounded subset ofJRnhas a compact closure. Since JR" is a prototype ofall finite-dimensional (normed) vector spaces, the same statement is true for all such vector spaces. What is interesting is that the statement indeed characterizes theDonned space: criterion for finite-dimensionality

16.4.11. Theorem. A normed vector space isfinite-dimensional if and only if every bounded subset has a compact closure. This result can also be applied to subspaces of a normed vector space: A subspace W of a normed vector space V is finite-dimensional if and ouly if every bounded subset ofW has a compact closure in W. A useful version of this property is stated in terms of sequences of points (vectors): 16.4.12. Theorem. A subspace W ofa normed vector space V is finite dimensional if and only if every bounded sequence in W has a convergent subsequence in W.

Karl Theodor WDhelm Weierstrass (1815-1897) was boththe greatest analyst andthe world's foremost teacher of advanced mathematics of the last third of the nineteenth century. His career was also remarkable in another way-and a consolation to all "late starters"-for he began the solid part of his professional life at the age of almost 40, whenmost mathematicians are long pasttheircreative years. Hisfather senthim to theUniversity of Bonn to qualify forthehigherranks of the Prussian civil serviceby studying law andcommerce. But Karl hadno interest in these subjects. He infuriated his fatherby rarely attendinglectnres, gettingpoor grades, and instead, becominga champion beerdrinker. He didmanage to becomea superb fencer, butwhenhe returned home, he hadno degree. In order to earn his living, he made a fresh start by teaching mathematics, physics, botany, German, penmanship. and gymnastics to the children of several small Prussian towns during the day. During the nights, however, he mingled with the intellectuals of thepast, particularly thegreat Norwegian mathematician Abel. Hisremarkable research on Abelian functions wascarried on foryearswithout theknowledge of another livingsoul;he didn'tdiscuss it withanyone atall, or submit it forpublication in themathernaticaljoumals of the day. All thischanged in 1854 whenWeierstrass at lastpublished anaccount of his research on Abelian functions. Thispaper caught theattention of an alert professor attheUniversity of Konigsberg who persuaded his university to award Weierstrass an honorary doctor's degree. The Ministry of Education granted Weierstrass a year'sleave of absence withpay

16.4 COMPACT SETS

463

to continue his research, andthe next yearhe was appointed to the University of Berlin, where he remained therestof his life. Weierstrass's great creative talents were evenly dividedbetweenhis thinking andhis teaching. The student notes of his lectures,andcopies of these notes, andcopies of copies, werepassed from hand to hand throughout Europe andevenAmerica. Like Gauss he was indifferent to fame, but unlike Gauss he endeared himself to generations of students by the generosity with whichhe encouraged themto develop andpublish, andreceivecredit for, ideas andtheorems that heessentially originated himself. Among Weierstrass's students andfollowers were Cantor, Schwarz, Holder, Mittag-Leffler, SonjaKovalevskaya (Weierstrass's favorite student), Hilbert, Max Planck, Willard Gibbs, andmanyothers. In 1885 he published the famous theorem now called theWeierstrassapproximation theorem (see Theorems 5.2.3 and 8.1.1), which was given a far-reaching generalization, withmanyapplications, by themodem American mathematician M. H. Stone. The quality thatcameto be knownas "Weierstrassian rigor" was particularly visible in his contributions to thefoundations ofreal analysis. He refused to acceptany statement as "intuitively obvious," butinstead demanded ironclad proofbasedon explicitproperties of therealnumbers. Thecareful reasoning required for theseproofswas founded on a crucial property of therealnumbers now knownas the BWHB theorem.

We shall need the following proposition in our study of compact operators: 16.4.13. Proposition. Let W be.a closed proper subspaceofX and 8 an arbitrary nonnegative number with 0 :0 8 < 1. Then there exists a unit vector Ivo} E X such that

v

IIx- voll2:8

[x) E W.

Proof Choose a vector Iv} in X but not in W and let

d = minl]» - xIII [x)

E

W}.

We claim that d > O. To show this, assume otherwise. Then for each (large) n and (sufficiently small) f, we could find distinct vectors (Ix n )) whose distance from [u) would be fin and for which the sequence (Ixn )) would have Iv} as a limit. Closure ofW would then imply that Iv} is in W, a contradiction. So, d > 0, Now, for any Ixo} E W, let eW

lu} sa [x) _ Iv} - Ixo} = '(IIv - xolllx) + Ixo}) -Iv} IIv -xoll IIv -xoll

and note that by the definition of d, the norm of the numerator is larger than d. Therefore, lIuli 2: dlllv - xoll for every Ix}, Ixo} E W. If we choose Ixo} snch that II v - Xo II < t, which is possible because d8- 1 > d, then lIu II 2: 8 for all Ix} E W. Now let Ivo} = (Iv) - Ixo})/lIv - xoll. 0

as:

464

16. AN INTROOUCTION TO OPERATOR THEORY

16.5 Compact Operators It is straightforward to show that if X is a compact set in 9i i and f : 9i 1 ..... 9i 2 is continuous, then fiX) (the image of X) is compact in 9i2. Since all bounded operators are continuous, we conclude that all bounded operators map compact subsets onto compact subsets. Thereis a specialsubsetof 21 (9i1, 9i2) that deserves

particular attention. compact operator

16.5.1. Definition. An operator K E 2l(9it, 9i2) is called a compact operator if it maps a bounded subset of9il onto a subset Of9i2 with compact closure.

Since we will be dealing with function spaces, and since it is easier to deal with sequencesof functionsthan with subsetsof the space of functions,we find it more useful to have a definition of compact operatorsin terms of sequencesrather than subsets.Thus,instead of a bounded subset, we take a subsetof it consistingof a (necessatily)bounded sequence.The image of this sequencewill be a sequence in a compact set, which, by defutition, must have a convergentsubsequence. We therefore have the following: 16.5.2. Theorem. An operator K E 2l(9it, 9i2) is compact ifand only iffor any bounded sequence (Ix n )} in 9it, the sequence (K Ixn )} has a convergent subsequence in 9i2. 16.5.3. Example. Consider 2l(:J{), thesetofbounded operators ontheHilbert space X, If K is a compactoperator and T a bounded operator, then KT and TK are compact. This is because (T IXn) == IYn)) is a bounded sequence if (Ix n)) is, and (K IYn) = KT IXn)) has a

product oftwo compact operators Is compact

convergent subsequence,because Kis compact. For the second part, use the first definition of the compact operator and note that K maps bounded sets onto compact sets, which T (being continuous) maps onto a compact set. As a special case of this property we note that the product of two compact operators is compact. Similarly, one can show that any linear combination of compactoperators is compact.Thus,anypolynomial of a compactoperator is compact. In particular,

~

n'

.

~

n'

.

(1-K)n = L...., .. ,(-K)] =1+ L...., ' . ,(-K)] ==1-Kn, j=O J .(n - J). j=l ] .(n - J).

where Kn is a compactoperator.

finite rank operators

16.5.4. Definition. An operator T E if its range is finite-dimensional.

III

.c (9i t, 9i2)

is called a finite rank operator

The following is clear from Theorem 16.4.12. 16.5.5. Proposition. A finite rank operator is compact In particular,every linear transformationof a finite-dimensional vector space is compact.

16.5 COMPACT OPERATORS

linear

transformations of finite-dimensional vector spaces are

compact

465

16.5.6. Theorem. IflKn} E ,c(Xj, X2) are compactandK E ,c(Xj, X2) is such that II K - Kn II --+ 0, then K is compact. Proof. Let {lxm)J be a bounded sequence in Xl. Let {Kl Ixm1)J be the convergent subsequence guaranteed by the compactness of Kj. Now, {lxm1)J is a bounded sequence in Xl. It therefore has a subsequence {Ix m,)} such that {K2Ixm,)} is convergent. Note that {Kj Ixm,)} is also convergent. Continuing this process, we

constroct the sequence of sequences

where each sequence is a subsequence of all sequences preceding it. Furthermore, all the sequences {I
"E/3trick." Write K IYm) - K IYn} = K IYm) -I
+ I
K1IYn)

+ I
K IYn}

and use the triangle inequality to obtain

By choosing m, n, and I large enough, we can make each of the three terms on the RHS smaller than E/3;the first and the third ones because Kl --+ K, the second one because {I
Kn sa

L

Cij Ie,) (ejl,

i,j=l

Clearly, 11K - Kn II --+ O. The hermitian adjoints {Kk} are also of finite rank (therefore, compact). Barring some convergence technicality, we see that Kt, which is the limit of the sequence of these compact operators, is also compact. Kis compact iff Kt is

16.5.7. Theorem. K is a compact operator if and only ifKt is. A particular type of operator occurs frequently in integral equation theory. These are called Hilbert-Schmidt operators and defined as follows:

Hiibert-Schmidt operators

16.5.8. Definition. Let:J{ be a Hilbert space, and {lei) }~l an orthonormal basis. An operator T E ,c(X) is called Hilberl-Schmidt if

466

16. AN INTROOUCTION TO OPERATOR THEORY

00

tr(TtT) sa

L (edTtT

00

lei) =

Hilbert-Schmidt kernel

00

Tei) =

i=l

i=l

Hilbert-Schmidt operators are compact.

L (Ted

L IITedl2 <

00.

i=l

16.5.9. Theorem. Hilbert-Schmidt operators are compact. For a proof, see [Rich 78, pp. 242-246]. 16.5.10. Example. It is time to give a concrete example of a compact (Hilbert-8chmidt) operator. For this, we retorn to Equation (16.2) with w(y) = I, and assume that Iu) E r}(a, b). Suppose further thatthe function K(x, y) is continuous on the closed rectangle [a, b] X [a, b] in the xy-plane (orIR2 ) . Under such conditions, K(x, y) is called a HilhertSchmidt kernel. We now show that K is compact. Firstnote thatdue to the continuity of K(x, y), IK(x, y)I 2dx dy < 00. Next, we calculate the trace of KtK. Let {lei) )~1 be any orthonormal basis of ,e2(a, b). Then

J: J:

trKtK =

f:

(etl KtK lei) =

l=l

=

fff

'tfff

(eil x) (xl Kt Iy) (yl Klz) (zl ei) dxdydz

/=1

(YIKIX)*(YIKlz)'t(zlei)(eilx)dXdydZ 1=1 ~o(x-z)

=

fff

(yl Klx)* (yl K Iz)'(zl

(~Iei) (eil) Ix) dxdydz

---=1

III

Bernard Balzano (1781-1848) was a Czech philosopher, math-

ematician, andtheologian whomade significant contributions to bothmathematics andthe theoryof knowledge. He enteredthe Philosophy Faculty .ofthe University of Prague in 1796, stodying

philosophy andmathematics. He wrote "Myspecialpleasure in mathematics restedtherefore particularly on its purelyspeculative parts, in otherwordsI prizedonly thatpart of mathematics which was at the same time philosophy." . 10 the automo of 1800 he began three years of theological studywhile he was preparing a doctoral thesis on geometry. He received his doctorate in 1804 for a thesis in which he gave his view of mathematics and what constitutes a correct mathematical proof. In the prefacehe wrote:

I could not be satisfiedwith a completely strictproof if it were not derived from concepts which the thesis to be provedcontained, butrather made use of some fortuitous, alien, intermediate concept,which is always an erroneous transition to another kind.

16.6 SPECTRUM OF COMPACT OPERATORS

467

Two days after receiving his doctorate Balzano was ordained a Roman Catholic priest.

However, he came to realize that teaching and not ministering defined his true vocation. In the same year, Balzano was appointed to the chair of philosophy and religion at the University of Prague. Because of his pacifist beliefs and his concern for economic justice, he was suspended from his position in 1819 after pressure from the Austrian government. Balzano had not given up without a fight but once he was suspended on a charge of heresy he was put under house arrest and forbidden to publish. Although some ofhis books had to be published outside Austria because of government censorship, he continued to write and to play an important role in the intellectual life of his country. Balzano intended to write a series of papers on the foundations of mathematics. He wrote two, the first of which was published. Instead of publishing the second one he decided to" . .. make myself better known to the learned world by publishing some papers which, by their titles, would be more suited to arouse attention." Pursuing this strategy he published Der binomische Lehrsatz ... (1816) and Rein analytischer Beweis ... (1817), which contain an attempt to free calculus from the concept of the infinitesimal. He is clear in his intention stating in the preface of the first that the work is "a sample of a new way of developing analysis." The paper gives a proof of the intermediate value theorem with Bolzano's new approach and in the work he defined what is now called a Cauchy sequence. The concept appears in Cauchy's work four years later but it is unlikely that Cauchy had read Balzano's work. After 1817, Bolzano published no further mathematical works for many years. Between the late I 820s and the 1840s, he worked on a major work Grossentehre. This attempt to put the whole of mathematics on a logical foundation was published in parts, while Balzano

hopedthat his students wouldfinish and poblishthe complete work. His work Paradoxien des Unendlichen, a study of paradoxes of the infinite, was published in 1851, three years after his death, by one of his students. The word "set" appears here for the first time. In this work Balzano gives examples of 1-1 correspondences between the elements of an infinite set and the elements of a proper subset. Bolzano's theories of mathematical infinity anticipated Georg Cantor's theory of infinite sets. It is also remarkable that he gave a function which is nowhere differentiable yet everywhere continuous.

16.6 Spectrum of Compact Operators Our next task is to investigate the spectrum 0" (K) of a compact operator K on a Hilbert space J£. We are particularly interested in the set of eigenvalues and eigenvectors of compact operators. Recall that every eigenvalue of an operator on a vector space of finite dimension is in its spectrum, and that every point of the spectrum is an eigenvalue (see page 457). In general, the second statement is not true. In fact, we saw that the right-shift operator had no eigenvalue at all, yet its spectrum was the entire unit disk of the complex plane. We first observe that 0 E O"(K), because otherwise 0 E p(K), which implies that K = K - 01 is invertible with inverse K- t. The product of two compact operators (in fact, the product of a compact and a bounded operator) is compact (see

468

16. AN INTRODUCTION TO OPERATOR THEORY

Example 16.5.3). This yields a contradiction? because the unit operator cannot be compact: It maps a bounded sequence to itself, not to a sequence with a couvergent subseqnence.

oE cr(K) if Kis compact

16.6.1. Proposition. For any compact operator K E sional Hilbert space, we have 0 E u(K).

~(:Ji)

on an infinite dimen-

To proceed, we note that eigenvectors of K corresponding to the eigenvalue ). belong to the null space of K - ),1. So, let lO

NJ.

= ker(K -

Nt

=ker(K t -),*1),

),1),

:RJ. es Range(K - ),1),

:Rt =Range(Kt -),*1).

16.6.2. Theorem. NJ. and Nt are finite-dimensional subspaces of X. Further_ Nt more, -n.L J\}.. )..0 Proof. We use Theorem 16.4.12. Let (Ix.)} be a bounded sequence in NJ.. Since K is compact, (K Ix.) = ).Ix.)} has a convergent subsequence. So (Ix.)} has a convergent subsequence. This subsequence will converge to a vector in NJ. if the latter is closed. But this follows from Proposition 16.4.6, continuity of K- ),1, the fact that NJ. is the inverse image of the zero vector, and the fact that any single point of a space, such as the zero vector, is a closed subset. Finite-dimensionality of Nt follows from the compactness of Kt and a similar argument as above. To show the second statement, we observe that for any bounded operator T, wehavelllu) ET(X).Liff(ulv) = o for all Iv) ET(X)iff(uITx) = o for all [x) E X iff (Ttul x) = 0 for all [x) EX iffTt lu) = 0 iff lu) E kerTt. This shows that T(X).L = ker rt, The desired result is obtained by letting T = K - ),1 and noting that (W.L).L = W for any subspace W of a Hilbert space. 0

We note that NJ. is the eigenspace of K corresponding to the eigenvalue ).. However, it may well happen that zero is the ouly number in o (K). In the finitedimensional case, this corresponds to the case where the matrix representation of the operator is not diagonalizable. In such a case, the standard procedure is to look at generalized eigenvectors. We do the same in the case of compact operators.

generalized eigenvector

16.6.3.Definition. A vector Iu) is a generaTked eigenvector of K of order m if (K - )'1)m- l l u ) # Obut(K- )'1)m lu) = O. The set ofsuch vectors, i.e., the null space of(K - )'1)m, will be denoted by Nt)

It is clear that {O}

=N(O)

C

N es N(ll C N(2) C ),.-).-

A-A

...

c_Jlfm) c N(m+l) A_A

...

l~Our conclusion is valid only in infinite dimensions. In finite dimensions, all operators, including 1, are compact. In what follows, we assume that X #= O. 11Recall that T(9{) is the range of the operatorT.

(16.6)

16.6 SPECTRUM OF COMPACT OPERATORS

469

and each Nt') is a subspace of 1C In general, a subspace with higher index is larger than those with lower index. If there happens to be an equality at one link of the above chain, then the equality continnes all the way to the right ad infinitum. To see this, let p be the first integer for which the equality occurs, and let n > p be arbitrary. Suppose lu) E Nin+l). Then (K - A1)p+1[(K - Al)n- p lu)] = (KA1)n+1 lu) = O. It follows that (K - Al)n- p lu) is in NjP+1). But NjP) = NjPH). So (K - At)" lu)

= (K -

Al)P[(K - Al)n- p lu)]

= o.

Thus every vector in Nin+1) is also in Nin). This fact and the above chain imply that Nin) = NinH) for all n > p, n

16.6.4. Theorem. The subspace ll Ni ) isfinite-dimensionalfor each n. Moreover, there is an integer p such that

for n = 0, 1,2, ... , p - 1 butNin ) = Nin+1) for all n :::: p, Proof For the first part, use the result of Exarnple 16.5.3 to show that (K-A t)" = Kn - An1 where Kn is compact. Now repeat the proof of Tbeorem 16.6.2 for Kn. If the integer p exists, the second part of the theorem follows from the discussion preceding the theorem. To show the existence of p, suppose, to the contrary, that Nin ) oft Nin+1) for every positive integer n. This means that for every n, we can find a (unit) vector Ivn) E Nin+l) that is not in Nin) and that by Proposition 16.4.13 has the property

We thus obtain a bounded sequence {Iv n)]. Let us apply K to this sequence. If j > I, then (Ivj) - IVi») E N~+I) by the construction of IVj) and the fact that N(l+l) c N}U.+1.) Furthermore , A -

but

11 Forinfinite dimensions, thefactthatlinearcombinations of a subsetbelong to the subsetis not sufficient to makethat subset into a subspace. The subsetmustalso be closed. We normally leave out therather technical proofof closure.

470

16. AN INTRODUCTION TO OPERATOR THEORY

by the definition ofN~+I). Therefore, (K - A1)(lvj) -Iv[)) E N~). Now note that ENf)

ENf)

,-'-,

,-"-,

Klvj) - Klvl) = AH
It follows from Proposition 16.4.13 that the nann of the vector in curly brackets is larger than ~. Hence, II K Ivj) - K IVI) II ::: IA I/2, i.e., since j and I are arbitrary, the sequence (K Ivn ) } does not have a convergent subsequence. This contradicts the fact that K is compact. So, there must exist a p such that Nip) = Nl."+!). D We also need the range of various powers of K- A1. Thus, let ~in) = Range(KAt)". One can show that 1C =

~(O) ::::> ~(l) ::::> ••• ::::> ~(n) ::::> ~(n+l) ::::> ••• ),.-),,-

-J..-)"

-

16.6.5. Theorem. Each ~in) is a subspace of 1C. Moreover, there is an integer q men) f=..AJ,. m(n+l)fiorn -- 0 , l men) -..AJ.. - m(n+l)fioraII n:::: q. sueh tha t J\}, , ...i , qb - , ut.}\,J. Proof The proof is similar to that of Theorem 16.6.4. The only extra step needed is to show that ~in) is closed. We shall not reproduce this step. D

16.6.6. Theorem. Let q be the integer ofTheorem 16.6.5. Then

1. 1C = Niq ) Ell ~iq) 2. Niq ) and ~iq) are invariant subspaces ofK.

3. The only vector in ~iq) that K - A1 maps to zero is the zero vector. Infact, when restricted to ~iq), the operator K - A1 is invertible. Proof (l) Recall that 1C = Niq ) Ell ~iq) means that every vector of 1C can be

written as the sum of a vector in Niq) and a vector in ~iq), and the only vector common to both subspaces in zero. We show the latter property first. In fact, we show that Nim) n ~iq) = 0 for any integer m. Suppose [x) is in this intersection. For each n ::: q, there must be a vector Ixn) in 1Csuch that [x) = (K - At)" Ixn) because~in) = ~iq) forn::: q. If [x) i' 0, then Ixn)
16.6 SPECTRUM OF COMPACT OPERATORS

471

From 0= (K-A1)'" [x) = (K-A1)'"+'lxr) and

it follows that Ixr) E Ni,"+r). But Ni,"+r) = Nf) = Nj"), contradicting Equation q), (16.7). We conclude that [x) must be zero, By the definition of:Ri for any vector q) q) lz) in :Ji, we have that (K - A1)q lz) E :Ri . Since :Ri = :Rfq), there must he a vector Iy) E:Ji such that (K-A1)q [z) = (K- A1)2q Iy) or (K- A1)qllz) - (Kq). A1)q Iy)] = O. This shows that Iz) - (K- A1)q Iy) is in Ni On the other hand, [z) = liz) - (K-A1)q ly)J+(K-A1)q Iy)..

and the first part of the theorem is done. (2) For the second part, we simply note that (K - A1)Nik ) C; Nik and that

q) K(Nl

= (K -

q) A1 + A1)(Ni

= (K -

q A1)(Ni

1 k ) C; Ni ) ,

)+ A1(Niq) c Niq).

c:Niq)

cNiq)

Similarly, q) K(:Ri

= (K -

q) A1 + A1)(:Ri

= (K - A1)(:Riq) + A1(:Riq) c:Riq+ c1l.i c:Riq )

q). C :Ri

~

1)

q) (3) Suppose [z) E :Ri and (K - A1) [z)

Q

)

= O. Then [z) = (K -

A1)q Iy) for some q+ Iy) in Jf, and 0 = (K - A1) [z) = (K - A1)q+lly), or Iy) E Ni 1) . From part (l)-withm = q+ I-we conclude that lz) = O. It follows that K-A1 is injective (or 1-1). We also have q) (K-A1):Ri = (K-A1)(K-A1)q(:Ji) q+ = (K - A1)q+l(:Ji) = :Ri 1) = :Riq). q), Therefore, when restricted to :Ri the operator K - A1 is surjective (or onto) as q) q) well. Thus, (K - A1) : :Ri --> :Ri is bijective, and therefore has an inverse. D 16.6.7. Corollary. The two integers p and q introduced in Theorems 16.6.4 and 16.6.5 are equal.

472

16. AN INTROOUCTION TO OPERATOR THEORY

Proof The proof is left as a problem for the reader (see Problem 16.5).

D

The next theorem characterizes the spectrum ofa compact operator completely. In order to prove it, we need the following lemma.

16.6.8. Lemma. Let K, : 9 9
1. Each nonzero point ofa(K) is an eigenvalue ofK whose eigenspace isfinitedimensional.

2. a(K,) "" a(K).

3. Every infinite sequence in a(K) converges to zero. Proof (I) If A oF 0 is not an eigenvalue of K, the null space of K - A1 is zero. This says that {O} "" NIO) "" Nil) "" .. " i.e., p "" q "" O. From Theorem 16.6.6, we conclude that:Ji "" NIO) EIl9
9
1. 0

E

a(K).

2. Each nonzero point ofa (K) is an eigenvalue ofK whose eigenspace isfinitedimensional.

3. a (K) is either a finite set or it is a sequence that converges to zero.

16.7 SPECTRAL THEOREM FOR COMPACT OPERATORS

473

o Figure 16.1 The shaded area represents a convexsubset of the vector space. It consists of vectors whose tips lie in the shaded region. It is clear that there is a (unique) vector belonging to the subsetwhose length is minimum.

Proof (I) was proved in Proposition 16.6.1. (2) was shown in the lemma above. (3) Let
16.7 Spectral Theorem for Compact Operators The finite-dimensional spectral decomposition theorem of Chapter 4 was based on the existence of eigenvalues, eigenspaces, and projection operators. Such existence was guaranteed by the existence of an inner product for any finite-dimensional vector space. The task of establishing spectral decomposition for infinite-dimensional vector spaces is complicated not only by the possibility of the absence of an inner product, but also by the questions of completeness, closure, and convergence. One can elintinate the first two hindrances by restticting oneself to a Hilbert space.

However, even so, onehasto dealwithother complications of infinite dimensions. As an example, consider the relation V = W 6) W.L, which is ttivially true for any subspace W in finite dimensions once an orthonormal basis is chosen. Recall that the procedure for establishing this relation is to complement a basis of W to produce a basis for the whole space. In an infinite-dimensional Hilbert space, we do not know a priori how to complement the basis of a subspace (which may be infinite-dimensional). Thus, one has to prove the existence of the orthogonal complement of a subspace. Without going into details, we sketch the proof. First a definition:

474

16. AN INTRODUCTION TO OPERATOR THEORY

convex subset

16.7.1. Definition. A convex subset E ofa vector space is a collection ofvectors such that iflu) and Iv) are in E, then lu) -t(lu) -Iv» is also in Efor all 0 :0: t :0: 1. Intuitively, any two points of a convex subset can be connected by a straigbt line segment lying entirely in the subset. Let E be a convex subset (not a subspace) of a Hilbert space Ji:. One can show that there exists a unique vector in E with minimal nonn (see Figure 16.1). Now let M be a subspace of Ji:. For an arbitrary vector lu) in Ji:, consider the subset E = lu) - M, i.e., all vectors of the form lu) - 1m) with 1m) E M. Denote the unique vector of minimal nonn of lu) - M by lu) - IPu) with IPu) EM. One can show that lu) - IPu) is orthogonal to lu), i.e., (Iu) - IPu» E M.L (see Figure 16.2). Obviously, only the zero vector can be simultaneously in M and M.L. Furthermore, any vector lu) in Ji: can be written as lu) = IPu) + (Iu) -IPu» with IPu) EM and (Iu) -IPu» E M.L. This shows that Ji: = M Ell M.L. In words, a Hilbert space is the direct Sum of anyone of its subspaces and the orthogonal complement of that subspace. The vector IPu) so constructed is the projection of lu) inM. A projection operator P can be defined as a linear operator with the property that p 2 = P. One can then show the following. 16.7.2. Theorem. The kernel ker P of a projection operator is the orthogonal complement ofthe range P(Ji:) of P in Ji: iff P is hermitian. This is the reason for demanding henniticity of the projection operators in our treatruent of the finite-dimensional case. We now concentrate on the compact operators, and first look at hermitian compact operators. We need two lemmas: 16.7.3. Lemma. Let H E :B(Ji:) be a bounded hermitian operator on the Hilbert space K Then [H] = maxl] (H»] x) II1Ixll = II.

Proof Let M denote the positive number on the RHS. From the definition of the norm of an operator, we easily obtain I (Hxlx) 1:0: lIHlllIx1l 2 = lIHlI,orM:o: lIHII. For the reverse inequality, see Problem 16.7.

D

16.7.4. Lemma. Let K E :B(Ji:) be a hermitian compact operator. Then there is an eigenvalue}" ofKsuch that IAI = lIKll.

Proof Let (Ixn)I be a sequence of unit vectors such lbat

This is always possible, as the following argument shows. Let E be a small positive number. There must exist a unit vector IXI) E Ji: such that

16.7 SPECTRAL THEOREM FOR COMPACT OPERATORS

475

Figure 16.2 The shaded area represents the subspace Jy( of thevectorspace. Theconvex subset E consists of all vectors connecting points of M to thetipof lu}. It is clearthat there is a (unique) vectorbelonging to E whose length is minimum. The figure shows that this vector is orthogonal to M,

because otherwise, IIKII - E would be greater thau or equal to the uonn of the operator (see Lemma 16.7.3). Similarly, there must exist auother (differeut) unit vector IX2} E 1f such that II KII - E12 = I (Kx21 X2) I. Continuing this way, we construct au infinite sequence of unit vectors (Ix n)) with the property IIKII-Eln = I (Kx n I x n) I·This coustruction clearly produces the desired sequence. Note that the argumeut holds for any hennitiau bounded operator; compactness is uot uecessary. Now define an sa (Kxnl x n ) aud leta = lim an, so that lal = IIKII. Compactness of Kimplies that (IKxn)) converges. Let Iy} E 1fbe the limit of {I Kxn}}. Then lIyll = lim IIKx nll ~ IIKllllxnll = IIKII. On the otherhaud,

o~

2. IIKxn - axnll = IIKx nll2 - 2a (Kxnl x n) + lal

Taking the limit aud noting that an aud a are real, we get

It follows from these two inequalities that lIyll Furthermore,

(K - al)(ly} la)

= (K -

al)(lim Ixn})

= IIKII aud that lim Ixn } = Iy} [tx,

= lim(K -

al) Ixn}

Therefore, a is au eigeuvalue of K with eigenvector Iy} I a.

=0 o

Let us order all the eigeuvalues ofTheorem 16.6.9 in decreasing absolute value. Let M n denote the (finite-dimensional) eigeuspace corresponding to eigenvalue An, aud Pn the projection to M n. The eigenspaces are pairwise orthogonal aud PnPm = 0 for m i' n. This follows in exact aualogy with the finite-dimensional

case. First assume that K has only finitely mauy eigenvalues,

476

16. AN INTRODUCTION TO OPERATOR THEORY

Let M "" Ml Ell Mz Ell ... Ell M,. = I:'i=l EIlMj, and let Mo be the orthogonal complementof M. Since each eigenspace is invariant underK, so is JvC, Therefore, by Theorem 4.2.3-which holds for finite- as well as infinite-dimensional vector spaces-and the fact that Kis henmitian, Mo is also invariant. Let Ko be the restriction of K to Mo. By Lemma 16.7.4, Ko has an eigenvalne A snch that IAI = IIKolI.· If A i" 0, it mnst be one of the eigenvalnes already accounted for, because any eigenvalue of Ko is also an eigenvalue of K. This is impossible, because Mo is orthogonal to all the eigenspaces. So, A = 0, or IAI = IIKoll = 0, or Ko = 0, i.e., K acts as the zero operator on Mo. Let Po be the orthogonal projection on Mo. Then JC = I:'i=o EIlMi- and we have 1 = I:'i=o Pi- and for an arbitrary [x) E JC, we have K [x) = K

(~Pj IX)) = ~K(Pj Ix)) = tAj(Pj Ix)).

It follows that K = I:'i=l AjPj. Notice that the range ofKis I:'i=l EIlMj, which is finite-dimensional. Thus, K has finite rank. Barring some technical details, which we shall notreproduce here, the case ofa compact henmitian operator with infinitely many eigenvalues goes through in the sanne way (see [DeVi 90, pp. 179-180]): spectral theorem for compact hermitian operators

16.7.5. Theorem. (spectral theorem: compact henmitian operators) Let K be a compact hermitian operator on a Hilbert space JC. Let {A j IN=l be the distinct nonzero eigenvalues of K arranged in decreasing order of ahsolute values. For each j let Mj be the eigenspace ofK corresponding to eigenvalue Aj and Pj its projection operator with the property Pi P j = 0 for i i" j. Then:

1. If N <

=

00,

then K is an operator of finite rank, K =

JC Mo Ell Ml Ell··· Ell MN, or 1 dimensional.

=

I: J= l AjPj, and

I:J=oPj, where Mo is infinite-

2. If N = 00, then Aj -+ 0 as j -+ 00, K = I:i=l AjPj, and JC = Mo Ell I:i=l EIlMl» or 1 = I:i=o P[, where Mo could be finite- or infinitedimensional. Furthermore,

which shows that the infinite series above converges for an operator norm. The eigenspaces of a compact hermitian operator are orthogonal and, by (2) of Theorem 16.7.5, span the entire space. By the Grann-Schrnidt process, one can select an orthonormal basis for each eigenspace. We therefore have the following corollary. 16.7.6. Corollary. If K is a compact hermitian operator on a Hilbert space JC, then the eigenvectors ofK constitute an orthonormal basis for JC.

477

16.7 SPECTRAL THEOREM FOR COMPACT OPERATORS

16.7.7. Theorem. Let K be a compact hermitian operator on a Hilbert space Jf and let K = '£7~1 AjPj, where N could be infinite. A bounded linear operator on Jf commutes with K if and only if it commutes with every PiProof The "if" part is straightforward. So assume that the bounded operator T commutes with K. For [x) E Mj,wehave(K-Aj)Tlx) = T(K-Aj) Ix) =0. Similarly, (K - Aj)Tt [x) = Tt(K - Aj) Ix) = 0, because 0 = [T, K]t = K]. These equatious show that both T aud rt leave M j invariaut. This meaus that M j reduces T, aud by Theorem 4.2.5, TPj = PjT. 0

rr'.

Next we prove the spectral theorem for a uormal operator. Recall that auy operatorTcaubewrittenasT=T,+iTiwhereT, = !(T+Tt)audTi = lICT-Tt) are henntitiau, aud since both T aud rt are compact, T, aud Ti are compact as well. For normal operators, we have the extra condition that [T" Til = [T,Tt] = O. Let T, = '£7~1 AjPj aud r, = ,£f=ll-'kQk be the spectral decompositions err, aud Ti. Using Theorem 16.7.7, it is strailfhtforward to show that if [T" Til = 0 then [Pj,Qk] = O.Now,sinceJf = '£j~oEllMj = '£~oEll:Nk.whereMj are the eigenspaces ofT, aud:Nk those ofTi, we have, for auy Ix) E Jf, T, [x) =

(~AjPj) (~Qk IX)) = ~~AjPjQk Ix).

Sintilarly, r, [x) = Ti('£7~oPj Ix) = '£f=1 ,£7=0I-'kQkPj [x), Combining these two relations aud noting that QkPj = Pj Qk gives N N T [x) = (T, + iTi) [x) = L L(Aj

+ il-'k)PjQk [z}.

j~Ok~O

The projection operators Pj Qk project onto the intersection of M j aud :Nk. Therefore, M j n :Nk are the eigenspaces of T. Only those terms in the sum for which M j n:Nk f= 0 contribute. As before, we cau order the eigenvalues according to their absolute values. spectral theorem for compact normal operators

16.7.8. Theorem. (spectral theorem: compact normal operators) LetT be a compact normal operator on a Hilbert space Jf. Let {A j 17=1 (where N can be 00) be the distinct nonzero eigenvalues ofT arranged in decreasing order ofabsolute values. For each n let M n be the eigenspace ofT corresponding to eigenvalue An and Pn its projection operator with the property PmPn = Ofor m f= n. Then: 1. If N < 00, then T is an operator of finite rank T = '£~~1 Aj Pi- and Jf = Mo Ell Ml Ell··· Ell MN, or 1 = ,£7=0 Pj, where Mo is infinitedimensional. 2. If N = 00, then An --> 0 as n --> 00, T = '£~1 AnPno and Jf = Mo Ell '£~1 EIlMn, or 1 = '£1=0 Pj, where Mo could be finite- or infinitedimensional.

478

16. AN INTRODUCTION TO OPERATOR THEORY

As in the case of a compact hermitian operator, by the Gram-Schmidt process, one can select an orthonormal basis for each eigenspace of a normal operator, in which case we have the following: 16.7.9. Corollary. lfT is a compact normal operator on a Hilbert space Jf, then the eigenvectors ofT constitute an orthonormal basis for Jf. One can use Theorem 16.7.8 to write any function of a normal operator T as an expansion in terms of the projection operators of T. First we note that T k has A~ as its expansion coefficients. Next, we add various powers of T in the form of a polynomial and conclude that the expansion coefficients for a polynomial p (T) are pO.n). Finally, for any function f(T) we have 00

f(T) = Lf(An)Pn.

(16.8)

n=l

Johann (John) von Neumann, (1903-1957), the eldest of three sons of Max von Neumann, a well-to-do Jewish banker, was privately educated until he entered the gymnasium in 1914. His unusual mathematical abilities soon came to the attention of his teachers, who pointed out to his father that teaching him conventional school mathematics would be a waste of time; he was therefore tutored in mathematics under the guidance of university professors, and by the age of nineteen he was already recognized as a professional mathematician and had published his first paper. Von Neumann was Privatdozent at Berlin from 1927 to 1929 and at Hamburg in 1929-1930, then went to Princeton.University forthree years; in 1933 he was invited to join the newly opened Institute for Advanced Study, of which he was the youngest permanent member at that time. At the outbreak of World War Il, von Neumann was called upon to participate in various scientific projects related to the war effort: In particular, from 1943 he was a consultant on the construction of the atomic bomb at Los Alamos. After the war he retained his membership on numerous government boards and committees, and in 1954 he became a member of the Atomic Energy Commission. His health began to fail in 1955, and he died of cancer two years later. It is only in comparison with the greatest mathematical geniuses of history that von Neumann's scope in pure mathematics may appear somewhat restricted; it was far beyond the range of most ofhis contemporaries, and his extraordinary work in applied mathematics, in which he certainly equals Gauss, Cauchy, or Poincare, more than compensates for its limitations. Von Neumann's work in pure mathematics was accomplished between 1925 and 1940, when he seemed to be advancing at a breathless speed on all fronts of logic and analysis at once, not to speak of mathematical physics. The dominant theme in von Neumann's work is by far his work on the spectraltheory of operators in Hilbert spaces. For twenty years he was the undisputed master in this area, which contains what is now considered his most profound and most original creation, the theory of rings of operators.

16.7 SPECTRAL THEOREM FOR COMPACT OPERATORS

479

The first papers (1927) in which Hilbert space theory appears are those on the foundations of quantum mechanics. These investigations later led von Neumann to a systematic study of unbounded hermitian operators. Von Neumann's most famous work in theoretical physics is his axiomatization ofquantum mechanics. When he began work in that field in 1927,' the methods used by its founders were hard to formulate in precise mathematical tenus: "Operators" on "functions" were handled without much consideration of their domain of definition or their topological properties, and it was blithely assumed that such "operators," when self-adjoint, could always

be "diagonalized" (as in the finite dimensional case), at the expense of introducing Dirac delta functions as "eigenvectors." Von Neumann showed that mathematical rigor could be restored by taking as basic axioms the assumptions that the states of a physical system were points of a Hilbert space and that the measurable quantities were Hermitian (generally unbounded) operators densely defined in that space. After 1927 von Neumann also devoted much effort to more specific problems of quantum mechanics, such as the problem of measurement and the foundation of quantum statistics and quantum thermodynamics, proving in particular an ergodic theorem for quantum systems. All this work was developed and expanded in Mathematische Grundlagen del" Quantenmechanik (1932), in which he also discussed the much-debated question of"causality" versus "indeterminacy" and concluded that no introduction of "hidden parameters" could keep the basic structure of quantum theory and restore "causality." Von Neumann's uncommon grasp of applied mathematics, treated as a whole without divorcing theory from experimental realization, was nowhere more apparent than in his work on computers. He became interested in numerical computations in connection with the need for quick estimates and approximate results that developed with the technology used for the war effort-particularly the complex problems of hydrodynamics-and the completely new problems presented by the harnessing of nuclear energy, for which no ready-made theoretical solutions were available. Von Neumann's extraordinary ability for rapid mental calculation was legendary. The story is told of a friend who brought him a simple kinematics problem. Two trains, a certain given distance apart, move toward each other at a given speed. A fly, initially on the windshield of one of the trains, flies back and forth between them, again at a known constant speed. When the trains collide, how far has the fly traveled? One way to solve the problem is to add up all the successively smaller distances in each individual flight. (The easy way is to multiply the Ily'sspeed by the time elapsed until the crash.) After a few seconds of thought, von Neumann quickly gave the correct answer. "That's strange," remarked his friend, "Most people try to sum the infinite series." "What's strange about that?" von Neumann replied. "That's what I did."

In closing this section, let ns remark that the paradigm of compact operators, namely the Hilbert-Schmidt operator, is such because it is defined on the finite rectangle [a, b) x [a, b). If this rectangle grows beyond limit, or equivalently, if the Hilbert space is r:.. 2 ( R oo) , where Roo is some infinite region of the real line, then the compactness property breaks down, as the following example illustrates. 16.7.10. Example. Consider the two kernels Kj(x. t) = e- Ix-I I

and

K2(X, t) = sinxt

480

16. AN INTROOUCTION TO OPERATOR THEORY

where the first one acts on ,(,2(_00, 00) and the second one on ,(,2(0, 00). One can show (see Problem 16.8) that these two kernels have, respectively, the two eigenfunctions e

ia t•

a E JR.,

a >0,

and

corresponding to thetwo eigenvalues

2 1+0<

),,=--2'

o
and

We see that in the first case, all real numbers between0 and2 are eigenvalues, rendering this set uncountable. In the second case, there are infinitely (in fact, uncountably) many eigenvectors. (oneforeach a) corresponding to thesingleeigenvalue JJr /2. Note, however, that in the first case the eigenfunctions and in the second case the kernel have infinite norms. II

16.8

Resolvents

The discussion of the preceding section showed that the spectrum of a normal compact operator is countable. Removing the compactness property in general will remove countability, as showninExample 16.7.10. We have also seen that the rightshift operator, a bounded operator, has uncountably many points in its spectrum. We therefore expect that the sums in Theorem 16.7.8 should be replaced by integrals in the spectral decomposition theorem for (noncompact) bounded operators. We shall not discuss the spectral theorem for general operators. However, one special class of noncompact operators is essential for the treatroent of Stnrrn-Liouville theory (to be stndied in Chapters 18 and 19). For these operators, the concept of resolvent will be used, which we develop in this section. This concept also makes a connection between the countable (algebraic) and the uncountable (analytic) cases. resolvent ofan operator

16.8.1. Definition. Let T be an operator and A E P (T). The operator Ri.(T) ea (T - A1)-1 is called the resolvent ofT at A. There are two important properties of the resolvent that are useful in analyzing the .spectrum of operators. Let us assume that A, fJ, E P(T), A rf fJ" and take the difference between their resolvents. Problem 16.9 shows how to obtain the following relation: (16.9) To obtain the second property of the resolvent, we formally (and indefinitely) differentiate Ri.(T) with respect to A and evaluate the result at A = W

16.8 RESOLVENTS

481

Differentiating both sides of this equation. we get 2Ri (T), and in general.

n

d -R,(T) = d)..n

n!R~+l(T)

n

d -R,(T) I

=}

,~"

d)..n

=

n!R~+l(T).

Assuming that the Taylor series expansion exists, we may write

(16.10) which is the second property of the resolvent. We now look into the spectral decomposition from an analytical viewpoint. For convenience, we concentrate on the finite-dimensional case and let A be an arbitrary (not necessarily hermitian) N x N matrix. Let Abe a complex number that is larger (in absolute value) than any of the eigenvalues of A. Since all operators on finite-dimensional vector spaces are compact, Lemma 16.7.4 assures us that 1)..1 > II TII. and it is then possible to expand R, (T) in a convergent power series as follows:

1 R,(A) = (A_)..l)-l = - -

A n L:(-) 00

).. n~O

(16.11)

)..

This is the Laurent expansion of R, (A). We can immediately read off the residue of R,(A) (the coefficient of 1/)..): Res[R,(A)] = -1

=}

----.!..., 1 2m

fr. R,(A)d)" =

1.

where I' is a circle with its center at the origin and a radius large enongh to encompass all the eigenvalues of A [see Figure l6.3(a)]. A similar argument shows that

----.!..., 1 )..R,(A)d)" 21n

fr.

= A.

and in general. - -1. 21l"1

i nn )..R,(A)d)..=A r

forn =0.1 •...

Using this and assuming that we can expand the function f(A) in a power series. we get

-~ 1 f()..)R,(A)d)" = 2m fr.

f(A).

Writing this equation in the form

_1_1 f()..) d).. = 2rri fr. )"l-A

f(A)

(16.12)

482

16. AN INTRODUCTION TO OPERATOR THEORY

0 0 0 0 0



• •



• •

0 (b)

(a)

Figure 16.3 (a) Thelargecircleencompassing all eigenvalues. (b) thedeformed contour consisting of smallcirclesorbiting the eigenvalues.

makes it recognizable as the generalization of the Cauchy integral formula to operator-valued functions. To use any of the above integral formulas, we must know the analytic behavior of RJ.(A). From the formula of the inverse of a matrix given in Chapter 3, we have

where Cjk(J..) is the cofactor of the ijth element of the matrix A - J.. 1 and p(J..) is the characteristic polynomial of A. Clearly, Cjk(J..) is also a polynomial. Thus, [RJ. (A)]jk is a rational function of J... It follows that RJ. (A) has ouly poles as singularities (see Example 10.2.2). The poles are simply the zeros of the denominator, i.e., the eigenvalues of A.We can deform the contour r in such a way that it consists of small circles Yj that encircle the isolated eigenvalues J.. j [see Figure 16.3(b)]. Then, with f(A) = 1, Equation (16.12) yields

Pj sa

-~ J 2n,

r"j

RJ.(A)dJ... (16.13)

It can be shown (see Example 16.8.2 below) that {Pj) is a set of orthogonal projection operators. Thus, Equation (16.13) is a resolution of identity, as specified in the spectral decomposition theorem in Chapter 4.

483

16.8 RESOLVENTS

16.8.2. Example. We want to show that the Pj are projection operators. First let i = j. Then12

P] =

1_.)2 1.

( __ 27l'Z

1. Rtt(A)dl-'. t; R)JA)dA jYi

Note that Aneed not be equal to 1-'. In fact, we are free to choose IA - Aj I > II-' - Aj I,i.e., let the circlecorresponding to ).. integration be outsidethatof JL integration.U Wecanthen rewrite the abovedouble integral as

P] =

(_f)2 ir1. (') ir1. V» 1rl

j

I )21. 1.

= ( __

21ri

1'r~},.) J;,~) I

=

Rie(A)Rtt(A)dAdl-'

j

I

[Rie(A) _ Rtt(A)] dAdl-'

).. -

JL

J... - JL

h~J {it,) Rie(A) dA i(.) A~I-' J

J

i<.) J

Rtt(A) du:

it,) J

~I-'} ,

A

wherewe used Equation (16.9) to go to the second line. Now note that and

i

dA

y~).) I

- - =2rri A - JL

because Alies outside rJIL) and n lies inside

P] = (-

yy). Hence,

2~i)2 {o- 'Ixi iiV» Rtt(A)dl-'} = - 2~i i y)Rtt(A)dl-' = Pj.

The remaining part, namely PjPk = 0 for k '" j, can be done sbnilarly (see Problem 16.10). .. Now we let I(A) = A in Eqnation (16.12), deform the contonr as above, and

write

(16.14) 12We have not discussed multipleintegrals of complexfunctions. A rigorous study of such integrals involves the theory of functions of several complexvariables-a subjectwe have to avoiddue to lack of space. However, in the simple case at hand, the theory of realmultipleintegrals is an honestguide. 13This is possiblebecausethepoles areisolated.

484

16. AN INTRODUCTION TO OPERATOR THEORY

It can be shown (see Problem 16.11) that OJ =

1. (A t;

Aj)" RA(A)

o:

In particular, since RA(A) has only poles as singularities, there exists a positive integer m such that OJ' = O. We have not yet made any assumptions about A. If we assume that A is hermitian, for example, theu RA(A) will have simple poles (see Problem 16.12). It follows that (A - Aj)RA(A) will be analytic at Aj for all j = 1,2,oo.,r,andOj =OiuEquatiou(16.14). We thus have

, A= I>jPj, )=1

which is the spectral decomposition discussed in Chapter 4. Problem 16.13 shows that the Pj are hermitian.

16.8.3. Example. The most general 2 x 2 hermitiao matrix is of the form A = (a~l a l2

a 12) , aZ2

where au and aZ2 are-real numbers. Thus,

whichhas roots Al = ![411

+ a22 -

)(all - 422)2 + 4141212].

A2 = ![all

+ a22 + )(all -

a22)2 + 4141212].

Theinverse of A- A1 canimmediately be written: I (a22 - A - 412 ) R(A)-(A-A1)- I 411-A A - det(A-A1) -ai2 I = (A - AI)(A - A2)

(aZ2 -

-ah

A

-a I 2 )

411 - A .

We want to verify that R). (A) has only simple poles. Two cases arise:

i' A2, then it is clear that RA(A) has simple poles. 2. If Al = A2, it appears that R}" (A) may havea pole of order 2. However, note thatif Al = A2. thenthe square roots in the above equationsmust vanish. This happens iff all = aZ2 sa a and aiz = O. It then follows that A} = A2 == a, and 1. If Al

RA(A)

I

= (A-a)2

(4-A 0

0)

a-A'

This clearly showsthat RA(A) has only simple poles in dtis case.

III

16.9 PROBLEMS

Jordan canonical form

485

If A is not hermitian, D j oF 0; however, D j is nevertheless nilpotent. That is, Dj' = 0 for some positive integer m. This property and Equation (16.14) can be used to show that A can be cast into a Jordan canonical form via a similarity transformation. That is, there exists an N x N matrix S such that

1

SAS- = J =

("'I

0 J2

0 0

0

0

...

n

...

where Jk is a matrix of the form I 0 0 0 A I 0 0 0 A I

0 0 0

0 0 0

0 0

A

I

A Jk =

0 0

in which A is one of the eigenvalnes of A. Different Jk may contain the same

eigenvalues of A. For a discussion of the Jordan canonical form of a matrix, see [Birk 77], [Denn 67], or [Halm 58].

16.9 Problems 16.1. Suppose that S is a bounded operator, T an invertible operator, and that

Show that S is invertible. Hint: Show that T- 1S is invertible. Thus, an operator that is "sufficiently close"to an invertible operator is invertible.

16.2. Let V and W be finite-dimensional vector spaces. Show that T E .G(V,W) is necessarily bounded. 16.3. Let Jf be a Hilbert space, and T E .G(Jf) an isometry, i.e., a linear operator that does not change the norm of any vector. Show that IITII = 1. 16.4. Show that (a) the unit operator is not compact, and that (b) the inverse of a compact operator cannot be bounded. Hint: For (b) use the resnlts of Example 16.5.3.

q

l

16.5. Prove Corollary 16.6.7. Hint: Let [x) E Ni + ) and write it as [x) = In)+ Ir) q) q). with In) E Ni and Ir) E :1
q).

Conclude that Ir)

= 0, Niq+l) = Niq) ,

486

16. AN INTROOUCTION TO OPERATOR THEORY

and q 2: p. To establish the reverse inequality, apply (K - I-.l)P to both sides of the direct sum of part (I) of Theorem 16.6.6, and notice that the LHS is ~i.P), the second term of the RHS is zero, and the first term is ~iq+p). Now conclude that p 2: q.

16.6. Let]«) E ~andletMbeasubspaceof1f.ShowthatthesubsetE = lu}-M is convex. Show that E is not necessarily a subspace of~. 16.7. Show that for any hermitian operator H, we have

+ y) - (H(x - Y)lx - y) + i[(H(x + iy)1 x + iy} - (H(x - iy)1 x -

4(Hxly} = (H(x +Y)lx

iy)].

Now let Ix) = I-.Iz) and Iy) = 1Hz) II-., where I-. = (IIHzll/llzlI)1/2, and show that IIHzll 2 = {Hx] y) ::: MllzlIlIHzlI, where M = maxl] (Hz] z) 1/IIIzIl 2}. Now conclude that IIHII ::: M.

16.8. Show that the twokemels Kj(x, r) = e- Ix- tl and K2(X, t) = sinxt, where the first one acts on ,(,2(-00,00) and the second one on ,(,2(0, 00), have the two eigenfunctions and

o

at

t

V2 e + a2 + t2 '

a> 0,

respectively, corresponding to the two eigenvalues

2 1-.=I+a 2 '

aEIR,

and

16.9. Derive Equation (16.9). Hint: Multiply R,,(T) by 1 = R"m(T - iLl) and R"m by 1 = R,,(T)(T - 1-.1). 16.10. Finish Example 16.8.2 by showing that PjPk = 0 for k

i' j,

16.11. Show thalD) = :FYi (I-. - I-.j)nR"(A)d): Hint: Use mathematical induction and the technique used in Example 16.8.2. 16.12. (a) Take the inner product of lu) = (A - 1-.1) [u) with Iv) and show that for a hermitian A, 1m (vi u) = -(1m 1-.) IIv 11 2. Now use the Schwarz inequality to obtain IIvll:::

I~~~I

=>

IIR,,(A)lu)lI:::

(b) Use this result to show that

I~~I~I'

16.9 PROBLEMS

487

e

where is the angle that A- Aj makes with the real axis and A is chosen to have an imaginary part. From this result conclude that R. (A) has a simple pole when A is hermitian.

16.13. (a) Show that when A is hermitian, [R.(A)]t = R•• (A). (b) Write A - Aj = rje ie in the definition of Pj in Equation (16.13). Take the hermitian conjugate of both sides and use (a) to show that Pj is hermitian. Hint: You will have to change the variable of integration a number of times.

Additional Readings 1. DeVito, C. Functional Analysis and Linear Operator Theory, AddisonWesley, 1990. Our treatment of compact operators follows this reference's discussion. 2. Glimm, J. and Jaffe, A. Quantum Physics, 2nd ed., Springer-Verlag, 1987. One of the most mathematical treatments ofthe subject, and therefore a good introduction to operator theory (see the appendix to Part I). 3. Reed, M. and Simon, B. FourierAnalysis, Self-Adjointness, Academic Press, 1980. 4. Richtmyer, R. Principles of Advanced Mathematical Physics, SpringerVerlag, 1978. Discusses resolvents in detail. 5. Zeidler, E. Applied Functional Analysis, Springer-Verlag, 1995.

17

-,------

_

Integral Equations

The beginning of Chapter 16 showed that to solve a vector-operator equation one transforms it into an equation involving a sum over a discrete index [the matrix equation of Equation (16.1 or an equation involving an integral over a continuous index [Equation (16.2)]. The latter is called an integral eqnation, which we shall investigate here using the machinery of Chapter 16.

n,

17.1 Classification Volterra and Fredholm equations offirstand second kind

Integral equations can be divided into two major groups. Those that have a variable limit of integration are called Volterra equations; those that have constant limits of integration are called Fredholm equations. If the unknown function appears only inside the integral, the integral equation is said to be of the first kind. Integral equations having the unknown function outside the integral as well as inside are said to be of the second kind. The four kinds of equations can be written as follows.

l' 1

K(x, t)u(t) dt = v(x),

Volterra equation of the Ist kind,

K(x, t)u(t) dt = v(x),

Fredhohn equation of the Ist kind,

b

u(x) = v(x)

l' +1

+

K(x, t)u(t)dt,

Volterra equation of the 2nd kind,

K(x, t)u(t)dt,

Fredhohn equation of the 2nd kind.

b

u(x) = v(x)

In all these equations, K (x, t) is called the kernel of the integral equation.

17.1 CLASSIFICATION

489

In the theory of integral equations of the second kind, one nsually multiplies the integral by a nonzero complex number A. Thus, the Fredhohn equation of the second kind becomes

u(x) = v(x)

1

+A

b

(17.1)

K(x, t)u(t) dt,

and for the Volterra equation of the second kind one obtains u(x) = v(x) characteristic value of an Integral equation

1

+A

X

(17.2)

K(x, t)u(t) dt.

A A that satisfies (17.2) with v(x) = 0 is called a characteristic vaiue of the integral equation. In the abstract operator language both equations are written as lu} = Iv) + AK lu}

=}

(K - A-I) lu)

=

-A -llv).

(17.3)

Thus A is a characteristic value for (17.1) if and only if A-1 is an eigenvalue of K. Recall that when the interval of integration (a, b) is finite, K(x, t) is called a Hilbert-Schmidt kemel. Example 16.5.10 showed that K is a compact operator, and by Theorem 16.6.9, the eigenvalues of K either form a finite set or a sequence that converges to zero. 17.1.1. Theorem. The characteristic values ofa Fredholm equation ofthe second kind either form a finite set or a sequence ofcomplex numbers increasing beyond limit in absolute value. Our main task in this chapter is to study methods of solving integral equations of the second kind. We treat the Volterra equation first because it is easier to solve. Let us introduce the notation

1 x

K[u](x)

==

and

K(x, t)u(t) dt

Kn[u](x) = K[K n- 1[u]](x) (17.4)

whereby K [u 1denotes a function whose value at x is given by the integral on the RHS of the first equation in (17.4). One can show with little difficulty that the associated operator K is compact. Let M = max {IK (x, t) I I a ::; t ::; x ::; b} and note that

1 x

IAK[u](x)1 =

IA

K(x, t)u(t)

dtl : ; IAIIMlliulloo(x - a),

where lIulioo == max{lu(x) I I x E (a, b)}. Using mathematical induction, one can show that (see Problem 17.1)

I(AK)n [u](x) I s 1'-lnlMlnllulioo

(x

a)n

,

n.

(17.5)

490

17. INTEGRAL EQUATIONS

Since b 2:: x, we can replace x with b and still satisfy the inequality. Then the inequality of Eqnation (17.5) will hold for all x, and we can write the equation as an operator norm ineqnality: II (J.K)n II :s: 1J.lnlMl n lIu II oo(b - a)n [n': Therefore,

and the series I:~o(J.K)n converges for all J.. In fact, a direct calculation shows that the series converges to the inverse of 1 - J.K. Thus, the latter is invertible and the spectrum of K has no nonzero points. We have just shown the following. Volterra equalion of the second kind has a unique solution and no nonzero characterislic value

17.1.2. Theorem. The Volterra equation ofthe second kind has no nonzero characteristic value. In particular, the operator 1 - J.K is invertible, and the Volterra equation of the second kind always has a unique solution given by the convergent infinite series u(x) = I:j:o J.i K i (x, t)v(t) dt where K i (x, t) is defined inductively in Equation (/7.4).

J:

VitoVolterra (1860-1940)wasonly t 1whenhe hecameinterested inrnathematics whilereading Legendre's Geometry. At theageof 13 . he began to study thethree bodyproblem andmade someprogress.

His familywereextremety poor (his fatherhad died whenVito was two years old) butafterattending lecturesat Florencehe was ableto proceedto Pisa in 1878. At Pisa he studiedunder Betti, graduating as a doctor of physics in 1882.His thesison hydrodynamics included someresults of Stokes, discovered later butindependently

by Volterra. He became Professor of Mechanics at Pisa in 1883, andupon Betti'sdeath, he occupied the chair of mathematical physics. Afterspending some time at Turin as the chair of mechanics, he was awarded the chair of mathematical physics at theUniversity of Romein 1900. Volterra conceived theideaof a theory of functions that depend on a continuous set of values of another function in 1883. Hadamard was later to introduce the word "functional," whichreplaced Volterra's original terminology. In 1890Volterrausedhisfunctional calculus toshowthat thetheory of Hamilton andJacobi fortheintegration of thedifferential equations of dynamics couldbe extended to other problems of mathematical physics. Hismostfamous work wasdoneonintegralequations.Hebegan thisstudy in 1884,and in 1896hepublished several papers on what is nowcalledtheVolterra integral equation. He continued to study functional analysis applications to integral equations producing a large number of papers on composition andpermutable functions. During the First World War Volterra joined the Air Force. He made many journeys to France andEngland to promote scientific collaboration. After the war he returned to theUniversity of Rome, and his interests moved to mathematical biology. He studied the Verhulst equation andthelogisticcurve. He also wrote on predator-prey equations. In 1922 Fascism seized Italy, andVolterra fought against it in the Italian Parliament. However, by 1930theParliament wasabolished, andwhenVolterra refused to take anoath

491

17.1 CLASSIFICATION

of allegiance to the Fascist government in 1931, he was forced to leave the University of Rome.From the followingyearhe lived mostly abroad, mainly in Paris, butalso in Spain andother countries.

17.1.3. Example. Differential equations canbe transformed into integral equations. For instance, considerthe SOLDE d 2u dx 2

du

+ Pt (x) dx + po(x)u =

ural =

r(x),

cj ,

u'(a) = C2.

By integrating the DE once, we obtain -du = -

dx

LX PI (t)u' (t) dt - LX po(t)u(t) dt + LX ret) dt + C2. a

a

a

Integrating the first integral by parts gives u'(x)

= -PI (x)u(x) + LX [p\ (t) -

po(t)]u(t) dt +

LX ret) dt +PI (a)ct + C2. ~g(X)

~f(X)

Integrating once moreyields

-LX PI (t)u(t) dt + LX f(s) ds + LX g(s) ds + (x = -LX PI (t)u(t) dt + LX ds l'[p\ (t) - po(t)]u(t) dt + LX ds l' ret) dt + (x - a)[PI (a)ct + C2] + ct = LX (x - t)[p\ (t) - poet)] - PI (t)} u(1) dt + LX (x - t)r(t) dt + (x - a)[PI (a)Cj + C2] + cr.

u(x) =

a)[PI (a)ct

+ C2]

(17.6)

where we haveused theformula

LX ds

l'

f(t) dt

= LX (x -

t)f(t) dt,

whichthereader mayverifyby interchanging theorder of integration on theLHS.Equation

(17.6) is a Volterra equation of thesecoud kind withkernel K(x, t) sa (x - t)[p\ (t) - poet)] - PI (t)

and vex) '" Neumann series

solution

f:

(x - t)r(t) dt

+ (x

- a)[PI (a)ct

+ C2] + ct.

III

We now outline a systematic approach to obtaining the infinite series of Theorem 17.1.2, which also works for the Fredholm equation of the second kind as we shall see in the next section. In the latter case, the series is guaranteed to converge

492

17. INTEGRAL EQUATIONS

only if IAIII KII < 1. This approach has the advantage that in each successive step, we obtain a better approximation to the solution. Writing the equation as Iu) = Iv)

+ AK Iu) ,

(17.7)

we can interpret it as follows. The difference between lu) and [u) is AK lu). If AK were absent, the two vectors lu) and Iv) would be equal. The effect of AK is to change lu) in such a way that when the result is added to [u), it gives lu). As our initial approximation, therefore, we take Iu) to be equal to Iv) and write luo) = [u), where the index reminds us of the order (in this case zeroth, because AK = 0) of the approximation. To find a better approximation, we always snbstitute the latest approximation for lu) in the RHS of Equation (17.7). At this stage, we have lUI) = Iv) + AK luo) = Iv) + AK Iv). Still a better approximation is achieved if we substitute this expression in (17.7):

The procedure is now clear. Once lun), the nth approximation, is obtained, we can get IUn+l) by substituting in the RHS of (17.7). Before continuing, let us write the above equations in integral form. In what follows, we shall concentrate on the Fredhohn equation. To obtain the result for the Volterra equation, one simply replaces b, the upper limit of integration, with x. The first approximation can be obtained by substituting v(t) for u(t) on the RHS of Equation (17.1). This yields

Uj(x) = V(X)+A

l

b K(x,t)v(t)dt.

Substituting this back in Equation (17.1) gives

l + l + l

U2(X) = v(x)

+A

= v(x)

A

= v(x)

A

b dsK(x, s)Uj(s) b dsK(x, s)v(s) b dtK(x, t)v(t)

+ A21bdt

[l

b K(x, s)K(s, t) dS] v(t)

+ A21b dtK 2(x, t)v(t),

J:

where K 2(x, t) sa K(x, s)K(s, t)ds. Similar expressions can be derived for U3(X), U4(X), and so forth. The integrals expressing various "powers" of K can be

obtained using Dirac notation and vectors with continuous indices, as discussed

493

17.1 CLASSIFICATION

in Chapter 6. Thus, for instance,

K

b

3(x,

b

t) sa (xl K(l IsIl (sll dSI)K ( l IS2) (s21 dS2) Kit}

n

Iu n } =

Iv} +).K [u) + ... + ).nKn Iv} = L().K)i Iv} ,

(17.8)

i~O

whose integral form is b

Un(x) = t).i l

(17.9)

Ki(x,t)v(t)dt. a

j=O

Here K i (x, t) is defined inductively by

KO(x, z) = (xl ~ It)

= (xI1It) = (xl t) = 8(x -

t),

b

K i (x, t) = (xl KKi- 1 It) = (xl K ( l Is) (si dS) Ki- I It} b

= l

K(x,s)Ki-l(s,t)ds.

The limit of Un (x) as n ->

00

gives

b

u(x) = f).i l j=O

Ki(x, t)v(t)dt.

(17.10)

a

The convergence of this series, called the Neumann series, is always guaranteed for the Volterra equation. For the Fredholm equation, we need to impose the extra condition 1).IIIKIl < 1.

o

17.1.4. Example. As an example, let us find the solution of u(x) = 1 +). f u(t) dt, a Volterra equation of the second kind. Here, v(x) = I and K(x, t) = 1, and it is straightforward to calculate approximations to u(x): Uo(x)

= v(x) = 1,

U2(X)

= 1 +A 10

{X

UI(X)

= 1 +A foX K(x,t)uo(t)dt = 1 +AX, .

(X

K(x,t)uI(t)dt = 1 +A 10 (I +M)dt

= I +AX +

A2X2 -2-'

494

17. INTEGRAL EQUATIONS

It is clear that the nth term will look like

u,,(x)

).,2x2

).}lXn

n

)jx i

2

n.

j=O

J.

= l+h+--+ ...+-,- = L-.,-.

As n --+ 00, we obtain u(x) = eAx • By direct substitution, it is readily checked that this is indeed a solution of the original integral equation. II

17.2 Fredholm Integral Equations We can use our knowledge of compact operators gained in the previous chapter to study Fredholm equations of the second kind. With A "" 0 a complex number, we consider the characteristic equation (l-AK)lu) =

Iv),

(17.11)

u(x) - AK[u](x) = v(x),

or

where all fuuctious are square-integrable on [a, bl, and K(x, t), the HilbertSchmidt kernel, is square-integrable on the rectangle [a, bl x [a, b]. Using Propositiou 16.2.9, we innnediately see that Equation (17.11) has a unique solutiou if IAIIiKIl < I, and the solutiou is of theform 00

lu)

= (1 - AK)-llv) = LA"K" Iv),

(17.12)

n=O

or u(x) = L~oAnK"[vl(x), where Kn[v](x) is defined as in Equation (17.4) except that now b replaces x as the upper limit of integration. 17.2.1. Example. Considerthe integral equation u(x) -

10

1

K(x, t)u(t) dt = x,

where

K(x, r) =

X

{t

ifO~x
ift
Here )" = 1; therefore, a Neumann series solution exists if IIKIf < 1. It is convenient to write K in terms of the theta function: 1

K(x, t) = xO(t - x)

+ to(x -

This gives IK(x, t)12 = x 20Ct - x) O(x - t)O(t - x) = O. Thus, we have

IIKII 2 =

f f dx

(17.13)

r),

+ t 20(x

- t) because 02(x - t)

=

O(x - t) aod

2 dtIK(x, t)1

= fo1dx folx20Ct-X)dt+ fo1dx fo 1t20(X-t)dt

[I

r

2

= fo dt fo x dx

[I

+ fo

r

dx f

o

2 t dt

[I

= fo

dt

(t3) [I (x 3) 1 "3 + fo dx 3" = 6'

1Recall that the theta function is defined to be 1 if its argument is positive, and 0 if it is negative.

495

17.2 FREDHOLM INTEGRAL EQUATIONS

Sincethisis less than I, theNeumann series converges, andwe have2 u(x) =

I» l

j=O

b Ki(x,t)v(t)dt=

flo

l

Ki(x,t)tdt= ffi(x).

j=O 0

a

j=O

The firstfew terms are evaluatedas follows:

fo(x) =

10

1

KO(x,t)tdt =

10

1

8(x,t)tdt =x

!J(x) = (I K(x,t)tdt= (I[X8(t-X)+t8(x-t)]tdt

. 10 =x

10

1 1

t dt

x

+ (X t2 dt = :: _ x

10

2

3 .

6

Thenext term is trickier than thefirst twobecause of theproduct of thetheta functions. We first substitute Equation (17.13) in the integral for the second-order term, and simplify h(x)

= 10

1

10

1

=

=x

K 2(x,t)tdt

= 10

1

1

t dt

10

10 [x8(s -

1

tdt

10

x)

tdt

10

1

K(x,s)K(s,t)ds

+ sO(x -

s)][sO(t - s)

1

sO(s-x)8(t-s)ds+x

10

1

+ tOrs -

t 2dt

10

t)] ds

1

O(s-x)8(s-t)ds

+ (\dt (I s20(x-s)8(t-s)ds+ (I t2dt (\O(X-s)O(s-t)ds.

10

10

.

10

10

Itis convenientto switch theorder of integration atthispoint, Thisis because of thepresence ofO(x - s) and O(s -x), which do not involve t and are best integrated last. Thus, we have h(x) =X

10

1

s8(s-x)ds /,1 tdt+x

10

1

8(s-x)ds

loS t 2dt St

+ foIS20(X-S)dS /,I t dt+ foIS8(X-S)dS fo

2dt

=x1IsdS(~_s2)+x11ds s3 + (Xs2dS(~_s2)+

x 2 2 5 1 3 1 5 =-x--x +-x 24 12 120

x

3 10

2

2

f'sds

10

s3

3

As a test of his/her knowledge of O-function manipulation, the reader is orged to perform

theintegration in reverse order. Adding all the terms, we obtain anapproximation for u (x) that is validfor 0::: x ::: 1: II 2Notethat in thiscase (Fredholm equation), we cancalculate the jth termin isolation. In theVolterra case, it was morenatural to calculate the solution up to a given order.

496

17. INTEGRAL EQUATIONS

We have seen that the Volterra equation of the second kind has a unique solution which can be written as an infinite series (see Theorem 17.1.2). The case of the Fredholm equation of the second kind is more complicated because of the existence of eigenvalues. The general solution of Equation (17.11) is discussed in the following: Fredholm alternative

17.2.2. Theorem. (Fredhohn Alternative) Let K be a Hilbert-Schmidt operator and A a complex number. Then either 1. A is a regular value of Equation (17.11)---or A-I is a regular point ofthe operator K-in which case the equation has the unique solution lu) = (1 AK)-I Iv), or 2. A is a characteristic value ofEquation (17.11) (A-1 is an eigenvalue ofthe operator K), in which case the equation has a solution if and only if Iv)

is in the orthogonal complement of the (finite-dimensional) null space of 1-A*Kt. Proof The first part is trivial if we recall that by definition, regular points of K are those complex numbers fl. which make the operator K - fl.1 invertible. For part (2), we first show that the null space of 1 - A*Kt is finite-dimensional. We note that 1 - AK is invertible if and only if its adjoint 1 - A*Kt is invertible, and A E P (K) iff 1.* E P (Kt). Since the specttum of an operator is composed of all points that are not regular, we conclude that A is in the specttum of K if and only if 1.* is in the specttum of Kt. For compact operators, all nonzero points of the specttum are eigenvalues. Therefore, the nonzero points of the specttum of Kt, a compact operator by Theorem 16.5.7, are all eigenvalues of Kt, and the nnll space of 1 - A*Kt is finite-dimensional (Theorem 16.6.2). Next, we note that the equation itself requires that Iv) be in the range of the operator 1 - AK, which, by Theorem 16.6.2, is the orthogonal complement of the nnll space of1 - A*Kt. D

Erik Ivar Fredholm (1866-1927) was born in Stockholm, the sonof a well-to-do merchant family. He received thebesteducationpossibleandsoon showedgreat promise in mathematics,

leaning especially toward the applied mathematics of practical mechanics in a year of study at Stockholm's Polytechnic

Institute. Fredholm finished his education atthe University of Uppsala, obtaining his doctorate in 1898. He also studied at

theUniversity of Stockholm during thissameperiodandeventuallyreceivedan appointment to the faculty there. Fredholm remained there therestof his professional life. His first contribution to mathematics was contained in his doctoral thesis, in whichhe studied a first-order partial differential equation in three variables, a problem thatarises in the deformation of anisotropic 'media. Several years later

497

17.2 FREDHOLM INTEGRAL EQUATlDNS

he completed this work by finding the fundamental solntion to a general elliptic partial differential equation withconstant coefficients. Fredholm is perhaps best knownfor his studies of the integral equation that bears his name. Such equations occur frequently in physics. Fredholm's genius led him to note the similarity betweenhis equation anda relatively familiar matrix-vector equation, resulting in his identification of a quantity that playsthe samerole in his equation as thedeterminant plays in thematrix-vector equation. He thusobtaineda methodfor determining theexistence of a solution and later used an analogous expression to derive a solution to his equation akinto theCramer's rulesolution to thematrix-vector equation. He further showedthat the solution could be expressed as a powerseries in a complexvariable. This latter resultwas considered important enoughthat Poincare assumed it without proof (in facthe wasunable to prove it) in a study of related partial differential equations. Fredholm then considered the homogeneous form of his equation. He showed that under certain conditions, the vectorspace of solutions is finite-dimensional. David Hilbert later extended Fredholm's workto a completeeigenvalue theory of theFredholm equation, whichultimately led to the discovery of Hilbert spaces.

17.2.1 Hermitian Kernel Of special interest are integral equations in which the kernel is hermitian, which occurs exactly when the operator is hermitian. Such a kernel has the property thar' (tl K Ix) or [K(x, t)]* K(t, x). For such kernels we can use the (r] K [r)" spectral theorem for compact hermitian operators to find a series solution for the integral equation. First we recall that

=

=

where we have used J.. j 1 to denote the eigenvalue of the operator" and expanded the projection operator in terms of orthonormal basis vectors of the corresponding finite-dimensional eigenspace. Recall that N can be infinity. Instead of the double sum, we can sum once over all the basis vectors and write K = L:~I J..;llu n) (unl· Here n counts all the orthonormal eigenvectors of the Hilbert space, and J..;1 is the eigenvalue corresponding to the eigenvector Iu n). Therefore, J..;I may be repeated in the sum. The action of K on a vector Iu} is given by 00

K lu) =

I>;I (unl u) lun).

(17.14)

11=1

3Sincewe aredealingmainlywithrealfunctions, hermiticity of Kimplies the symmetry of K, i.e., K (x, t) = K (t, x). 4)" j is the characteristic valueof theintegral equation, orthe inverse of theeigenvalue of thecorresponding operator.

498

17. INTEGRAL EQUATIONS

If the Hilbert space is .(,2[0, b], we may be interested in the functional form of this equation. We obtain such a form by multiplying both sides by (x I: 00

K[u](x)

00

== (xIKlu) = I);! (unlu)

(xlu n)

= I);l (unlu)un(x).

n=l

Hilbert-Schmidt theorem

n=l

'

That this series converges uniformly in the interval [0, b] is known as the HilbertSchmidt theorem.

I:

17.2.3. Example. Let us solve u(x) = x + A K(x, t)u(t)dt, where K(x, t) '" xt is a symmetric (hermitian) kernel, by theNeumann series method. Wenote that /

ll =l l l b

IIKII 2 =

b

IK(x,t)1

b

b

x2dx

or

b

IIKII

=

2dxdt

t2dt =

ll b

=

(l

b

b x

2t2dxdt

2

X2dX) =

~(b3 _ a 3)2,

x 2dx = !(b 3 - a 3) ,

andtheNeumann seriesconverges if I).. I (b3 - a 3) < 3. Assuming that thiscondition holds; we have

\

00 u(x) = x + I>i

j=l

lb

Ki(x, t)tdt.

a

ThespecIal formof thekernel allows us to calculate xt (x, t) directly:

l t: =l xsfs~ 1 r b

Ki(x,t) =

K(x,SI)K(s!,S2) ... ·K(si_J,t)dslds2···dsi_l

b lb···lb

a

a

= xt (

a

b s2 ds

...

S~_ltdslds2··· dSj_l J

= xtllKlli- 1

I:

Ki (x, t)t dt = xIlKlli- 1!(b 3 - a 3 ) = xllKlli. Substituting this in the expression for u(x) yields

It follows that

u(x)

00

00

j=l

j=l

= x + I>ixlIKlli = x +xAIiKII I>i-11IKll i- 1 =X

(I +AIIKIII_~IIKII) = I-:IIKII

3x

Becauseof thesimplicity of thekernel, we cansolve theintegral equation exactly. First we write

l

u(x) = x + A

b xtu(t) dt

=x + AX

l

b tu(t) dt sa x(1 + AA),

(17.15)

499

17.2 FREDHDLM INTEGRAL EQUATIDNS

where A = A

f:

tu(t) dt. Multiplying both sidesby x aud integrating, we obtain

= lab xu(x)dx = (I +AA) lab x 2dx = (I + AA) II KII

=}

A

= I ~~::KII

Substitutiug A in Equation(17.15)gives

_ (

A IIKII

u(x) - x 1+ 1- AIIKII

)_ x - 1- AIIKII"

Thissolution is thesameas thefirst onewe obtained. However, no serieswasinvolved here, and therefore no assumption is necessary concerning l'AIIiKIl. II

If one cau calculate the eigenvectors Iun} and the eigenvalues A;;1, then one can obtain a solution for the integral equation in terms of these eigenfunctions as follows: Substitute (17.14) in the Fredholm equation [Equation (17.3)] to get 00

lu) = Iv)

+ A ~:::>;;1 (unl u) Iun).

(17.16)

n=l

Multiply both sides by (u mI: 00

(uml u) = (uml v)

+ ALA;;1 (unl u) (uml un) =

(uml v)

~

n=1

+ ).).;;;1

(uml U), (17.17)

or,if ).. is not one of theeigenvalues.

( Urn I) U

- Am(umlv) -

Am - A

.

Substituting this in Equation (17.16) gives lu} = Iv)

~ (unl v) _ A Iun),

+ AL. A n=l

(17.18)

n

and in the functional form, u(x) = v(x)

~ (unl v) + AL. -r-r--r-un (x), n=l

(17.19)

An - A

In case A = Am for some m, the Fredholm alternative (Theorem 17.2.2) says that we will have a solution only if Iv) is in the orthogonal complement of the null space of 5 1 - AmK. Moreover, Equation (17.17) shows that (uml u), the expansion coefficients of the basis vectors of the eigenspace M m , cannot be specified. 5Remember thatKis hermitian; therefore, Am is real.

500

17. INTEGRAL EQUATIONS

However, Equation (17.17) does determine the rest of the coefficients as before. In this case, the solution can be written as

~ (k) ~ + L.., ckl u", } + A L..,

lu) = Iv)

k=l

(u,,1 v) A _ A lu,,),

11=1

(17.20)

n

n=f:.m

where r is the (finite) dimension of M"" k labels the ottbononnal basis {Iu!:»} of M"" and {Ck}k~1 are arbitrary constants. In functional form, this equation becomes u(x) = v(x)

r

00

(ulv)

k=l

11=1

An - A

+ I: CkU~~)(X) + A I: _n_un(x).

n"",

(17.21)

17.2.4. Example. We now give an example of the application of Equation (17.19). We want to solve u(x) = 3 J~1 K(x, t)u(t) dt + x 2 where

and Pk(x) is a Legendrepolynontial. We first note that {Uk} is an orthonormal set of functions, that K(x, t) is real and symmetric (therefore, hermitian), and that

1 1

1

1 dt

-1

1 1 t Uk(X~U:(t) 1 1 = I:

dxIK(x, t)1 2 =

-1

1

1

dt

dx

-1 00

-1

k,/=O

1 I k/2172

k,/=O 2

2

u/(x;u;(t) Z/

2 /

1

-I

1

Uk(X)U/(x) dx =Okl

001

=

-1

Uk(t)U/(t) dt

~~~~

=clkl

001

I: "k 8kk = k=O I: "kZ = 2 <

k~O Z

00.

Thus, K (x, t) is a Hilbert-Schutidt kernel. Now note that

This ahows that zq is an eigenfunctionof K(x, r) witheigenvalue1/2k/ 2. Since3 ,. I/Z k/ 2 for any integer k, we canuse Equation (17.19) to write 2

u(x) = x

+ 3 I: 00

k=O

J~1 Uk(S)s2ds k/2 Z - 3

Uk(X).

17.2 FREDHDLM INTEGRAL EQUATIDNS

501

But J~I Uk(S)s2ds = 0 for k 2:: 3. For k ::: 2, we use the first three Legendrepolynomials to get 1

1

-I

..ti

2

uo(s)s ds = - ,

3

1

1 UI (s)s2ds

= 0,

-I

!-

This gives u(x) = 2x 2. The reader is urged to substitute this solution in the original integral equationandverify that it works. II

17.2.2 Degenerate Kernels degenerate or separable kernel

The preceding example involves the simplest kind of degenerate, or separable, kernels. A kernel is called degenerate, or separable, if it cao be written as a finite sum of products of functions of one variable: n

(17.22)

K(x, t) = I>/>j(x)1frj(t), j=l

where
lb

n

u(x) - A ~
If we define f.' j =

1: 1frj(t)u(t) dt, the preceding equation becomes

n

(17.23)

u(x) - A Lf.'j
Multiply this equation by 1frt(x) aod integrate over x to get n f.li -

A LfJ-jAij

= Vi

for i = I, 2, ... , n,

(17.24)

j~l

where Aij =

J:

J:

1frt(t)V(t) dt. With f.'i, Vi, aod Aij as

(1-AA)u =v.

(17.25)

1frt(t)
components of column vectors u, v, and a matrix A,we can writethe above linear system of equations as u - )"Au = v,

or

We cao now determine the f.'i by solving the system of linear equations given by (i7.24). Once the f.'i are determined, Equation (17.23) gives u(x). Thus, for a degenerate kernel the Fredholm problem reduces to a system of linear equations.

502

17. INTEGRAL EQUATIONS

17.2.5. Example. As a concrete example of anintegral equation withdegenerate kernel,

=

=

we solve u(x)-l fJ(I+xt)u(t) dt x for two different values ofA. Thekemel, K(x, t) 1 + xt, isseparable with 9'>1 (x) = I, 1/Jl (I) = I, 9'>2(X) = x, and 1/J2(t) = t. This gives the matrix

A=

G t)· 2

3

Forconvenience,we definethe matrix B sa 1 - )"A. (a) Firstassumethat)" = 1. In thatcase Bhas a nonzero determinant. Thus,8- 1 exists, and can be calculatedto be

With VI

=

f

1/J;(I)v(t)dt

=

f

tdt

=!

and

"2

= 10

1 1/J; (I )V(t ) d t

= 10

1 t

2dt

=j

we obtain ( fl.I ) = B-Iv = fl.2

(-i -2

Equation (17.23) then gives u(x) = fl.19'>1 (x) + f1.2¢2(x) + x = -2. (b) Now, for the purpose of illustrating the otheralternative of Theorem 17.2.2, let us take l = 8 + 203. Then

7 + 203 + 03

B = 1 - AA = - ( 4

(5

4 + 03

+ 203)/3

)

'

anddet B = O. This shows that8 + 2.../13 is a characteristic value of the equation. We thus have a solution only if vex) == x is orthogonal to the null space of 1 - A*At = at. To determine a basisforthisnull space, we have to find vectors Iz) suchthat at lz) = O. Since ). is real, andB is real andsymmetric, Bt = B, andwe must solve

(74++203 03

4

+ 03 ) (5 + 203)/3

«I) <2 = O.

The solution to this equation is a multiple of lz) == (-2!,JI3 ).lfthe integral equation is to have a solution,the columnvector v (whose corresponding ket we denote by Iv)) must be orthogonal to lz). But (zl v) = (3

-2 - 03)

(t) f-

O.

Therefore, the integral equationhas no solution.

III

503

17.2 FREDHOLM INTEGRAL EQUATIONS

The reader may feel uneasy that the functions
Alb

K(x, t)u(t) dt,

f:

we obtain u(x) = AL:?=l hi (x) h~(t)u(t) dt. Defining I-'i sa and substituting it back in the equation gives

f: h~(t)u(t)

dt

n

u(x) = A I)i(x)l-'i.

(17.26)

i=l

Multiplying this equation by A-lh;;(x) and integrating over x yields

This is an eigenvalue equation for the hermitian n x n matrix Mwith elements mi], which, by spectral theorem for hermitian operators, can be solved. In fact, the matrix need not be hetmitian; as long as it is normal, the eigenvalue problem can be solved. Once the eigenvectors and the eigenvalues are found, we can substitute them in Equation (17.26) and obtain u (x). We expect to find a finite number of eigenfunctions and eigenvalues. Our analysis of compact operators included such a case. That analysis also showed that the entire (infinite-dimensional) Hilbert space could be written as the direct sum of eigenspaces that are finite-dimensional for nonzero eigenvalues. Therefore, we expect the eigenspace corresponding to the zero eigenvalue (or infinite characteristic value) to be infinite-dimensional. The following example illustrates these points. 17.2.6. Example. Let us find thenonzero characteristic values andcorresponding eigen+ t) for -Jl' ::0 X. t ::0 n,

functions of the kernel K(x, t) = 1 + sin(x

6Actually, the problem of a degenerate kernel that leads to a normal matrix, as described below,canalso be solved.

504

17. INTEGRAL EQUATIONS

i:

We are seeking functions u and scalars A satisfying u(x) = AK[u](x), or

u(x) = A

Expanding sin(x u(x)

=)"

[I

+ sin(x + t)lu(t) dt.

+ t), we obtain

L:

[1 + sinx cos t

+ cos x sint]u(t) dt,

(17.27)

or

A-lu(x) = ILl

+ IL2 sinx + JL3 COSX,

(17.28)

where 1'1 = f':.n u(t) dt, !'2 = f':.n u(t) cost dt, and !'3 = f':.n u(t) sin t dt. Integrate both sides of Equation (17.28) with respect to x from -rr to 1C to obtain A-I JL 1 = 23l' JL 1. Similarly, multiplying by sin x and cosx andintegrating yields (17.29)

and

If 1'1 i' 0, we get A-I

= 2,.., which, when substituted in (17.29), yields!'2 = 1'3 = O. We

thus have, as a first solution, All = 2Jr and lUI)

=

a(6), where a is an arbitrary constant.

o

Equation (17.28) now gives Ai1uI(X) = 1'10 or UI(X) = cj , where

Cj

is an arbitrary

constant to be determined. On the other hand, 1'1 = 0 in-I i' 2,... Then Equation (17.29) yields A-I 1 =,.., Equation (17.28) gives 1'2 = ±1'3· For A-I ==

1.+

= ±,.. and

u(x) ea u+(x) = c+(sinx + cosx),

andfor A-I == ).,=1 = -Jr, it yields u(x) ss u_(x) = c_{sinx - COS x), where c± are arbitrary constants to be determined by normalization of eigenfunctions. The normalized eigenfunctions are I

U1=

$'

u±(x) =

~(sinx ± cos x).

v 2,..

Direct substitution in the original integral equation easily verifies that U I,'U+, and u.: are eigenfunctions of the integral equation withthe eigenvalues calculated above. Let us now consider the zeroeigenvalue (orinfinite characteristic value). Divideboth sides of Equation (17.27) by A and take the limit of A -+

00.

Then the integral equation

becomes

L:

[1 + sinx cost + cos x sint]u(t) dt = O.

The solutions u(t) to this equation would spanthe eigenspace corresponding to the zero eigenvalue, or infinite characteristic value. We pointed out above that this eigenspace is expected to be infinite-dimensional. This. expectation is borne out once we note that all functions of the form sin nt or cos nt with n ~ 2 makethe above integral zero; and there are infinitely many suchfunctions. III

17.3 PROBLEMS

505

17.3 Problems 17.1. Use mathematical induction to derive Equation (17.5). 17.2. Repeat part (a) of Example 17.2.5 using ¢2(x)

so that we still have K(x, t) =

= x,

'h(t)

=t

4>1 (X)1/tl(t) + ¢2(X)1/r2(t).

17.3. Use the spectral theorem for compact hermitian operators to show that if the kernel of a Hilbert-Schmidt operator has a finite number of nonzero eigenvalues, then the kernel is separable. Hint: See the discussion at the beginning of Section 17.2.1. 17.4. Use the method of successive approximations to solve the Volterra equation u(x) = A u(t)dt. Then derive a DE equivalent to the Volterra equation (make sure to include the initial condition), and solve it.

g

17.5. Regard the Fourier transform,

1

00

F[f](x) sa -I-

,J2ir

e

ixy

fey) dy

-00

as an integral operator. (a) Show that F2[f](x) = fe-x). (b) Deduce, therefore. that the only eigenvalues of this operator are A = ± I, ±i. (c) Let f(x) be any even function of x. Show that an appropriate choice of a can make u = f + aF[f] an eigenfunction of F. (This shows that the eigenvalues of F have infinite multiplicity.) 17.6. For what values of A does the following integral equation have a solution? u(x) = A

10" sin(x + t)u(t) dt + x.

What is that solution? Redo the problem using a Neumann series expansion. Under

whatcondition is the seriesconvergent? 17.7. It is possible to mnltiply the functions 4>j(x) by rj(x) and 1/rj(t) by I!yj(t) and still get the same degenerate kernel, K(x, t) = I:j=1 4>j (x)1/rj(t). Show that such arbitrariness, although affecting the matrices A and S, does not change the solution of the Fredholm problem u(x) -

Alb

K(x, t)u(t)dt = f(x).

17.8. Show, by direct substitution, that the solution found in Example 17.2.4 does satisfy its integral equation.

506

17. INTEGRAL EQUATIONS

17.9. Solve u(x) =

! f~1 (x + t)u(t) dt + x.

17.10. Solve u(x) = ),. f~ xtu(t) dt + x using the Neumann series method. For . what values of)" is the series convergent? Now find the eigenvalues and eigenfunctions of the kernel and solve the problem using these eigenvalues and eigenfunctions. 17.11. Solve u(x) = ),. f ooo K(x, t)u(t)dt+x a, where", is any real number except a negative integer, and K(x, t) = e-(x+I). For what values of)" does the integral equation have a solution? 17.12. Solve the integral equations

fol xtu(t)dt. (c) u(x) = x 2 + fol xtu(t) dt. (a) u(x) = eX +),.

(b) u(x)

=),.

(d) u(x) = x

fo" sin(x - t)u(t) dt.

+ foX u(t) dt.

17.13. Solve the integral equation u(x) = x +),. f~(x +t)tu(t) dt, keeping terms upto),.2. 17.14. Solve the integral equation u(x) = e- ixi +),. f.""oo e-1x-11u(t) dt, assuming that f remains finite as x --> ±oo.

oo 17.15. Solve the integral equation u(x) = e- ixi +),. fo u(t) cosxt dt, assuming that f remains finite as x --> ±oo.

Additional Reading 1. DeVito, C. Functional Analysis and Linear Operator Theory, AddisonWesley, 1990. 2. Jiirgen, K. Linear Integral Operators, Pitman, 1982. Translated from its original German, this is a thorough (but formal) introduction to integral operators and equations. 3. Zeidler, E. Applied Functional Analysis, Springer-Verlag, 1995.

18,

_

Sturm-Liouville Systems: Formalism

The linear operators discussed in the last two chapters were exclusively integral operators. Most applications of physical interest, however, involve differential operators (DO). Unfortunately, differential operators are unbounded. We noted that complications arise when one abandons the compactness property of the operator,

e.g., sums tum into integrals and one loses one's grip over the eigenvalues of noncompact operators. The transition to unbounded operators further complicates malters. Fortunately, the formalism of one type of DOs that occur most frequently in physics can be stndied in the context of compact operators. Such a stndy is our aim for this chapter.

18.1 Unbounded Operators with Compact Resolvent As was pointed out in Example 16.2.7, the derivative operator cannot be defined for all functions in .(,2(a, b). This motivates the following:

domain ofa linear operator

18.1.1. Definition. Let D be a linear manifold 1 in the Hilbert space 11:. A linear map T : :D --> 11: will be called a linear operator in2 11:. :D is called the domain ofT and often denoted by :D(T). 18.1.2. Example. The domainof the derivative operator D, as an operatoron £,2(a, b), cannot be the entire space. On the other hand, D is defined on the linear manifold M in £,2(a. b) spanned by (ei2nrrx/Lj wiih L = b - a. As we saw in Cbapter 8, :M is dense

1A linear manifold of an infinite-dimensional normedvector space '\7 is a proper subset that is a vector space in its own right, butls not necessarilyclosed. As opposed to on X.

508

18. STURM-LIOUVILLE SYSTEMS: FORMALISM

(see Definition 16.4.5 and the discussion following it) in ~2(a, b). This is the essence of

Fourier series: That every function in t.}(a, b) canbe expanded in (i.e., approximated by) a Fourier series. It turns out that many unbounded operators on a Hilbertspace sharethe same property, namely that their domains aredense in theHilbert space. Another important property of Fourier expansion is the fact that if the function is differentiable, then one can differentiate both sides, i.e., one can differentiate a Fourier expansion term by term if such anoperation makes sense for theoriginal function. Define the sequence (fm) by m

fm(x) =

L

uni2rrnx/L,

a'l =

1 .JL

l

b . a !(x)e-l271:nxjLdx.

n=-m

Then we can state the property above as follows: Soppose (fm) is in M.lflim [m = I and lim I:" = g, then I' = g and I E M. Manyunboundedoperatorsshare thisproperty. iii 18.1.3. Definition. Let :D be a linear manifold in the Hilbert space Jf. Let T :

:D --'> Jf be a linear operator in Jf. Suppose that for any sequence {Iun)) in D, both {Iun)) and {T lun)} converge in Jf, i.e., lim lun ) = lu) closed operator

and

lim T lun) = Iv) .

We say thatT is closed iflv) E :DandTlu) = [u). Notice that we cannot demand that Iv) be in :D for a general operator. This, as we saw in the preceding example, will not be appropriate for unbounded operators. The restriction of the domain of an unbounded operator is necessitated by the fact that the action of the operator on a vector in the Hilbert space in general takes that vector out of the space. The following theorem (see [DeVi 90, pp. 251-252] for a proot) shows why this is necessary: 18.1.4. Theorem. A closed linear operator in Jf that is defined at every point of Jf (so that :D = Jf) is bounded. Thus, if we are interested in unbounded operators (for instance, differential operators), we have to restrict their domains. In particular, we have to accept the possibility of an operator whose adjoint has a different domain.f

difference between hermitian and self-adjoint operators

18.1.5. Definition. Let T be a linear operator in Jf. We shall say thatT is hermitian ifTt is an extension ofT, i.e., :D(T) C :D(Tt) andTt lu) = T lu)foralllu) E :D(T). T is called self-adjoint if:D(T) = :D(Tt). As we shall see shortly, certain types of Sturm-Liouville operators, although unbounded, lend themselves to a study within the context of compact operators.

operators with compact resolvent

18.1.6. Definition. A hermitian linear operator T in a Hilbert space Jf is said to have a compact resolvent if there is a I-' E P (T) for which the resolvent R" (T) is compact.

3Thissubtledifference betweenhermitian andself-adjoint is statedheremerelyto warnthereader andwill be confined to the present discussion. The two qualifiers will be (ab)used interchangeably in therestof the book.

18.1 UNBOUNOEO OPERATORS WITH COMPACT RESOLVENT

509

An immediate consequence of this definition is that R, (T) is compact for all

pen.

AE To see this, note that R,(T) is bounded by Definition 16.3.1. Now use Equation (16.9) and write

The RHS is a product of a bounded" and a compact operator, and therefore must be compact. The compactness of the resolvent characterizes its spectrum by Theorem 16.7.8. As the following theorem shows, this in tnrn characterizes the spectmm of the operators with compact resolvent. 18.1.7. Theorem. Let T be an operator with compact resolvent R,(T) where A E pm. Then 0 'I I" E p(R,m) if and only if (A + 1/1") E peT). Similarly, I" 'lOis an eigenvalue ofR,(T) ifand only if (A + 1/1") is an eigenvalue ofT. Furthermore, the eigenvectors ofR,(T) corresponding to I" coincide with those of T corresponding to (A + 1/1").

Proof The proof consists of a series of two-sided implications involving definitions. We give the proof of the first part, the second part being very similar: I" E P (R, (T))

iff

R,(T) -1"1

is invertible.

R,(T) -1"1

is invertible iff

(T - A1)-1 -1"1

is invertible.

(T - Al)-1 -1"1

is invertible iff

l-I"(T-Al)

is invertible.

l-I"(T-Al)

is invertible iff

-l-T+Al I"

(±H)l-T

is invertible iff

(± H) E peT).

I

is invertible.

Comparing the LHS of the first line with the RHS of the last line, we obtain the first part of the theorem. D A consequence of this theorem is that the eigenspaces of an (unbounded) operator with compact resolvent are finite-dimensional, i.e., such an operator has only finitely many eigenvectors corresponding to each ofits eigenvalues. Moreover, arranging the eigenvalues I"n of the resolvent in decreasing order (as done in Theorem 16.7.8), we conclude that the eigenvalues of T can be arranged in a sequence in increasing order of their absolute values and the limit of this sequence is infinity.

r

18.1.8. Example. Consider the operator T in .(,2(0, 1) defined byS Tf = having the domain 2J(T) = If E .(,2(0, I) I E .(,2(0, I), frO) = f(l) = OJ. The reader may

r

4rhe sumof two bounded operators is bounded. 5Weshalldepart from ourconvention hereand shallnotuse theDirac bar-ket notation although theuse of abstract operators encourages theiruse. The reason is thatin this example, we aredealingwith functions, andit is moreconvenient to undress the functions fromtheirDiracclothing.

510

18. STURM-LIOUVILLE SYSTEMS: FORMALISM

check that zero is not an eigenvalue of T. Therefore, we may choose Ro(n = T- 1. We shall studya systematic way of finding inverses of some specific differential operators in theupcoming chapters on Green's functions. At this point,sufficeit to say that T- 1 canbe written as a Hilbert-Schmidt integral operator with kernel K(x,t) = {X(I-t) (I - x)t

Thus,ifT 1= g, i.e., if I" I(x)

= K[g](x) =

f

if 0 ~ x ~ t if 0 ~ t ~ x

s ~

I, I.

= -g, then rig = I, or 1= K[g], i.e., K(x, t)g(t) dt

= fox (I -

x)tg(t) dt

+ {(I -

x)tg(t) dt.

lt is readilyverified that K[g](O) = K[g](l) = 0 and !,,(x) = K[g]"(X) = -g. We can now use Theorem 18.1.7 with)" = 0 to find all the eigenvalues ofT: /Ln is an eigenvalue of T if andonly if 1/ Jkn is aneigenvalue of T- 1. These eigenvalues shouldhave finite-dimensional eigenspaces, andwe shouldbe ableto arrange them in increasing order

of magnitude without bound. To verify this, we solve I" = -J.L!. whose solutions are /Ln = n27f2 and In(x) = sinmf x. Note that there is only one eigenfunction corresponding to each eigenvalue. Therefore, theeigenspaces arefinite- (one-) dimensional. l1li

Sturm-Liouville operators

The example above is a special case of a large class of DOs occurring in mathematical physics. Recall from Theorem 13.5.4 that all linear second-order differential equations can be made self-adjoint. Moreover, Example 13.4.12 showed that any SOLDE can be transformed into a form in which the first-derivative term is absent. By dividing the DE by the coefficient of the second-derivative term if necessary, the study of the most general second-order linear differential operators boils down to that of the so-called Sturm-Liouville (S-L) operators d2 Lx == dx 2

separated boundary conditions

- q(x),

which are assumed to be self-adjoint. Differential operators are necessarily accompanied by boundary conditions that specify their domains. So, to be complete, let us assume that the DO in Equation (18.1) acts on the subset of £,z(a, b) consisting of functions u that satisfy the following so-called separated boundary conditions:

+ fhu'(a) =

0, azu(b) + fJzu'(b) = 0,

alu(a)

regular Sturm-Liouville systems

(18.1)

(18.2)

where aI, az, fh, and fJz are real constants with the property that the matrix of coefficients has no zero rows. The collection ofthe DO and the boundary conditions above is called a regular Sturm-Liouville system. We now show that the DO of a regular Sturm-Liouville system has compact resolvent. First observe that by adding au-with a an arbitrary number different from all eigenvalues of the DO-to both sides of the eigenvalue equation u" -

18.1 UNBOUNDED OPERATORS WITH COMPACT RESOLVENT

511

= AU, we can assume" thatzero is not an eigenvalue of Lx. Next, suppose that Ut(x) and uz(x) are ihe two linearly independent solutions ofihe homogeneous

qu

DE satisfying ihe first and ihe second boundary conditions of Equation (18.2), respectively. The operator whose kernel is K(x, t) = {-U1(X)UZ(t)/w(a) -UI(t)UZ(x)/W(a)

if a::o x ::0 t::o b, if a::o t::o x ::0 b,

in which W is ihe Wronskian of the solutions, is a Hilbert-Schmidt operator and iherefore compact. We now show ihat K(x, t) is ihe resolvent Ro(Lx ) = L.;-t sa K of our DO. To see this, write Lxu = v, and u(x)

= K[v](x) = -uz(x) -W(a)

1

l

x

b

Ut(x) Ut(t)v(t)dt - -(-) uz(t)v(t)dt. a Wax

Differentiating this once gives t

uz_ (x) u' (x) = __

1 x

u';t(x) UI (t)v(t) dt - _ _

W(a) a

W(a)

l

b

uz(t)v(t) dt,

x

and a second differentiation yields

1

l

x b u"(x) u"(x) u"(x) = __ z_ uI(t)v(t)dt __ t_ uz(t)v(t)dt W(a) a W(a) x

+ v(x).

The last equation follows from ihe fact ihat the Wronskian ui Uz- u;u I is constant for a DE of the form u" -qu = O. By substituting ul = qu; andu~ = qu: in ihe last equation, we verify ihat u = K[v] is indeed a solution ofihe Sturm-Liouvi1le system Lxu = v. Next, we show that ihe eigensolutions of ihe S-L system are nondegenerate, i.e., ihe eigenspaces are one-dimensional. Suppose II and [z are any two eigenfunctions corresponding to ihe same eigenvalue. Then boih must satisfy ihe same DE and the Same boundary conditions; in particular, we must have

0, '*

at!t(a) + fhI{(a) = at!z(a) + fhI~(a) = 0

(!t(a) jz(a)

I{(a)) (at) _ I~(a) fJt -

(0)

0 .

(18.3)

If at and fJI are not boih zero, ihe Wronskian-ihe determinant of ihe matrix above-must vanish. Therefore, ihe two functions must be linearly dependent. Finally, recall ihat a Hilbert space on which a compact operator K is defined can be written as a direct sum of ihe latter's eigenspaces. More specifically, 9{ = '£-7=0 EIlJV( j, where each JV(j is finite-dimensional for j = 1, 2, ... , and 6Although this will change q-and the originaloperator-no information will be lost because the eigenvectors will be the sameand all eigenvalues will be changed by a.

512

lB. STURM-LIOUVILLE SYSTEMS: FORMALISM

N can befinite or infinite. If N is finite, then Mo, which can be considered as the eigenspace of zero eigenvalne,7 will be infinite-dimensional. If Mo is finitedimensional (or absent), then N mnst be infinite, and the eigenvectors of K will span the entire space, i.e., they will form a complete orthogonal system. We now show that this holds for the regular Sturm-Liouville operator.

Jacques Charles Fraucois Sturm (1803-1855) madethefirst accurate determination of the velocity of sound in water in 1826, working with the Swiss engineer Daniel Colladon. He became a French citizen in 1833 and worked in Paris at the EcolePolytechnique where he became a professor in 1838. In 1840 he succeeded Poisson in the chair of mechanics in the Pacuue des Sciences, Paris. The problems of determining the eigenvalues andeigenfunctions of an ordinary differential equation with boundary conditions and of expanding a given function in tenus of an infinite series of the eigenfunctions, which date from about 1750, becamemoreprominent as new coordinate systems were introduced andnew classes of functions aroseas the eigenfunctions of ordinary differential equations. Sturm andhis friend Joseph Liouville decided to tackle the general problem for any second-order linear differential equation. . Sturm hadbeenworking since 1833on problems of partial differential equations, primarily on the flow of heatin a bar of variable density, andhence was fully aware of the eigenvalue andeigenfunction problem. Themathematical ideas he applied to thisproblem arecloselyrelated tohisinvestigations ofthe reality anddistribution of theroots of algebraic equations. His ideas on differential equations, he says, came from the study of difference equations anda passage to thelimit. Liouville, informed by Sturm of theproblems he was working on, took up the same subject. The results of their joint work was published in several papers which are quite detailed.

Suppose that the above Hilbert-Schmidt operator K has a zero eigenvalue. Then, there must exists a nonzero function v such that K[v](x) = 0, i.e., -

1

l

x b U2(X) Ut(x) Uj (t)v(t) dt - - U2(t)V(t) dt = 0 W(a) a W(a) x

--

(18.4)

for all x. Differentiate this twice to get U"(x) - .Ji(a)

r

Ja

Ut (t)v(t) dt -

ur

u"(x)

~(a)

l

b x U2(t)V(t) dt

+ v(x) =

O.

Now substitute = qUj and U = qU2 in this equation and use Equation (18.4) to conclude that v = O. This is impossible because no eigenvector can be zero. 7Thereader recallsthat when Kactson Mo, it yields zero.

z

16.2 STURM--L1DUVILLE SYSTEMS AND SDLDES

513

Hence, zero is not an eigenvalue of K, i.e., Mo = {D). Since eigenvectors of K = L; 1coincidewith eigenvectors of Lx, andeigenvalues of Lx arethereciprocals of the eigenvalues of K, we have the following result. Theorem forregular Sturm-Liouville systems

18.1.9. Theorem. A regular Sturm-Liouville system has a countable number of eigenvalues that can be arranged in an increasing sequence that has infinity as its limit. The eigenvectors ofthe Sturm-Liouville operator are nondegenerate and constitute a complete orthogonal set. Furthermore, the eigenfunction Un (x) corresponding to the eigenvalue An has exactly n zeros in its interval ofdefinition. The last statement is not a result of operator theory, but can be derived using the theory of differential equations. We shall not present the,details of its derivation. We need to emphasize that the boundary conditions are an integral part of SL systems. Changing the boundary conditions so that, for example, they are no longer separated may destroy the regularity of the SoL system.

18.2 Sturm-Liouville Systems and SOLDEs We are now ready to combine our discussion of the preceding section with the knowledge gained from our study of differential equations. We saw in Chapter 12 that the separation of PDEs normally results in expressions of the form

L[u] + AU = 0,

d 2u P2(x)d 2 x

or

du

+ PI(x)-d + po(x)u + AU = x

0, (18.5)

where u is a function of a single variable and Ais, a priori, an arbitrary constant. This is an eigenvalue equation for the operator L, which is not, in general, self-adjoint. lfwe use Theorem 13.5.4 and multiply (18.5) by

1 w(x) = --exp . P2(X)

[fX -PI(t) - d t], P2(t)

it becomes self-adjoint for real A, and can be written as

d [ p(x) dx dU] dx

Liouville substitution

+ [AW(X) -

(18.6)

q(x)]u = 0

withp(x) = w(x)p2(x)andq(x) = -po(x)W(x). Equation (18.6) is the standard form of the SoL equation. However, it is not in the form studied in the previous section. To turn it into that form one changes both the independent and dependent variables via the so-called Lionville substitntion: u(x) = v(t)[p(x)w(x)r 1/4,

_lX~(S)d . s.

t-

a

pes)

(18.7)

514

18. STURM-LIOUVILLE SYSTEMS: FORMALISM

It is then a mailer ofchain-rule differentiation to show that Equation (18.6) hecomes

d2v dt 2

+ [J. -

(18.8)

Q(t)]v = 0,

where Q(t) = q(x(t)) w(x(t))

+ [P(X(t))W(X(t))]-1/4 d: [(pw)I/4]. dt

Therefore, Theorem 18.1.9 still holds.

Joseph Liouville (1809-1882) was a highly respecled professor at the College de France, in Paris, and the founder and editor of

the Journal des Mathbnatiques Pures et Appliquees, a famousperiodical that played an important role in French mathematical life through the latter part of the nineteenth century. His own remark-

able achievements as a creativemathematicianhave only recently received the appreciation they deserve. He was the first to solve a boundary value problem by solving an equivalent integral equation. His ingenious theory of fractional differentiation answered the long-standing question of what reasonable meaning can be assigned to the symbol an y / dx n when n is not a positive integer. He discovered the fundamental result in complex analysis that a bounded entire function is necessarily a constant and used it as the basis for his own theory of elliptic functions. There is also a well-known Liouvilletheoremin Hamiltonian mechanics, which states that volume integrals are time-invariant in phase space. In collaboration with Sturm, he also investigated the eigenvalue problem of second-order differential equations. The theory of transcendental numbers is another branch of mathematics that originated in Liouville's work. The irrationality of 1f and e (the fact that they are not solutions of any linear equations) had been proved in the eighteenth century by Lambert and Euler. In 1844 Liouville showed that e is not a root of any quadratic equation with integral coefficients as well. This led him to conjecture that e is transcendental, which means that it does not satisfy any polynomial equation with integral coefficients.

18.2.1. Example. The Liouville substitutiou [Equatiou (18.7)1 transformsthe Bessel DE (xu')' + (Px - v2/x)u = 0 iuto 2

d 2v -+

[2

dt

2

1/4] v=O,

k - v -2 t

from which we can obtain aninteresting result when v =

!.In that case we have v+k2v = 0,

whose solutious are of the form coskt aud siukt. Noting that u(x) = JI/2(X), Equatiou (18.7) gives

sinkt JI/2(kt) = A .,ji

or

coskt Jl/2(kt) = B .,ji ,

18.2 STURM--LIOUVILLESYSTEMS ANO SOLDES

515

and since h/2(X) is analyticatx = 0, wernnst have h/2(kt) = A sinkt/,f!, whichis the resultobtainedin Chapter 14.

..

The appearance of W is the result of our desire to render the differential nperator self-adjoint. It also appears in another context. Recall the Lagrange identity for a . self-adjoint differential operator L: d , , uL[v] - vL[u] = -{p(x)[u(x)v (x) - v(x)u (x)]). dx

(18.9)

If we specialize this identity to the S-Lequation of(18.6) with u = Uj corresponding to the eigenvalue Al and v = U2 corresponding to the eigenvalue A2, we obtain for the LHS uIL[U2] - u2L[uIl = Uj(-A2WU2)

+ U20.IWUI) =

(AI - A2)WUIU2.

Integrating both sides of (18.9) then yields (Aj - A2) lab WUjU2dx =

periodic boundary conditions

{p(x)[Uj(x)u~(x) - U2(X)U~ (x)])~.

(18.10)

A desired property of the solutions of a self-adjoint DE is their orthogonality when they belong to different eigenvalues. This property will be satisfied if we assume an inner product integral with weight function W (x), and if the RHS of Equation (18.10) vanishes. There are various boundary conditions that fulfill the latter requirement. For example, Uj and U2 could satisfy the boundary conditions of Equation (18.2). Another set of appropriate boundary conditions (BC) is the periodic BC given by ural = u(b)

and

u' (a) = u' (b).

(18.11)

However, as the following example shows, the latter BCs do not lead to a regular S-L system. 18.2.2. Example. (a)The S-Lsysternconsisting of the S-Lequationd 2u/dt2 +",2u = 0 in theinterval[0, T] withthe separatedBCs u(O) = 0 and u(T) = ohas theeigenfunctions un(t) = sin n; t with n = 1,2, ... and the eigenvalues An = w~ = (nx /T)2 with n = 1,2, .... (b) Let the S-Leqnation be the same as in part (a) bntchange the intervalto [-T, +T] and theBCs to aperiodic one suchas u(-T) = u(T) andu'(-T) = u'(T). The eigenvalues are the sameasbefore, but the eigenfunctions are 1, sin(mrt/T), andcos(mrt/T), wheren is a positive integer. Note that there is a degeneracy here in the sense thattherearetwo linearly independent eigenfunctions having the sameeigenvalue (mr / T)2. By Theorem18.1.9, the S-L system is not regular. . (c) The Besselequation for a given fixed 1)2 is 2

1 + ( k2 - x v ) u = 0, u" + -;.u' 2

where a :S:X :s:b,

516

18. STURM-LIOUVILLE SYSTEMS: FORMALISM

andit canbe turned intoanS-L system if we multiply it by I w(x)=--exp P2(X)

[IX --dt=exp Pl(t)] [IX -dt]= x . P2(t)

t

Thenwe can write

2 d (dU) x - + (2 k x - -v ) u = 0, dx x

dx

which is in theform of Equation (18.6) with P = w = x, A = k 2, and q(x) = v 2/x. If a > 0, we can obtain a regular SoL systemhy applyingappropriateseparatedBes. III singular S·L systems

A regular SoL system is too restrictive for applications where either a or b or both may be infinite or where either a or b may be a singular point of the SoL equation. A singular S·L system is one for which one or more of the following conditions hold: 1. The interval [a, b] stretches to infinity in either or both directions. 2. Either p or w vanishes at one or both end points a and b. 3. The function q(x) is not continuous in [a, b]. 4. Anyone of the functions ptx), q(x), and w(x) is singular at a or b. Even though the conclusions concenting eigenvalues of a regular SoL system cannot be generalized to the singular SoL system, the orthogonality of eigenfunctions correspondiug to different eigenvalues can, as long as the eigenfunctions are square-integrable with weight function w(x):

18.2.3. Box. The eigenfunctions ofa singular SoL system are orthogonal if the RHS of(18.10) vanishes.

18.2.4. Example. Bessel functions Jvex) are entire functions. Thus, they are squareintegrable in the interval[0, b] for any finitepositive b. For fixed v the DE

2 d2u r dr 2

2 2 + r du dr + (k r -

v

2

)u = 0

(18.12)

transforms into the Bessel equationx-e" +xu' +(x 1_v2)u = oifwe make the substitution kr = x. Thus, the solution of the singular S-L equation (18.12) that is analytic at r = 0 and corresponds to theeigenvalue k 2 is uk(r) = JvCkr). Fortwo different eigenvalues,

kr

and k~, the eigenfunctions are orthogonalif the boundaryterm of (18.10)corresponding to Equation(18.12)vanishes, that is, if (r[Jv(klr)J~(k2r)- Jv(k2r)J~(klr)]Jg

18.3 OTHER PROPERTIES OF STURM--L10UVILLE SYSTEMS

517

vanishes,whichwilloccurif and onlyif Jv(kj b)J~(k2b)-Jv(k2b)J~(kjb) = O. Acornmon choice is to take Jv(k]b) = 0 = Jv(k2b), that is, to take both k]b and k2b as (different) roots of the Bessel function of order v. Wethushave r Jv(kir)Jv(kjr) dr = 0 if ki and kj aredifferent roots of Jp(kb) = O. TheLegendre equation

It

2 )dU] -d [ ( I - x - +AU=O, dx dx

where - 1 < x < 1,

is aheady self-adjoint. Thus, w(x) = I, and p(x) = 1 - x 2. The eigenfnnctions of this singularS-I. system[singularbecanse p(I) = p(-I) = 0] are regular at the endpoints x = ±I and are the Legendrepolynomials Pn(x) correspondingto A = n(n + 1). The bonndary term of(I8.1O) clearlyvanishesata = -1 and b = +1. Since Pn(x) are square-integrable on [-I, +1], we obtain the familiar orthogonality relation: J~l Pn(x)Pm(x)dx = 0 if m

i' n. The Hermite DE is u" - 2xu' + AU

(18.13)

= O.

x2

It is transformedinto an SoL system if we multiply it by w(x) = e- . The resulting SoL equation is -d [ e- x2 _dU] dx dx

+ Ae-x2 u = O.

(18.14)

The boundary tenn corresponding to the two eigenfunctions u 1(x) and U2 (x) having the respective eigenvalues A1 and)..2 =1= Al is (e-

X2

[U] (x)u2 (x)- U2 (x)u! (x)]J~.

This vanishes for arbitrary Uland andb =

U2

(because they areHermite polynomials) if a = -00

+00.

The function u is an eigenfunction of (18.14) corresponding to the eigenvalue x if and only if it is a solution of (18.13). Solutions of this DE corresponding to A = 2n are the Hermite polynomials Hn (x) discussed in Chapter 7. We can therefore write J~:: e-x 2 Hn(x)Hm(x) dx = 0 ifm i' n. This orthogonalityrelation was also derivedin Chapter7. III

18.3 Other Properties of Sturm-Liouville Systems The S-L problem is central to the solution of many DEs in mathematical physics. In some cases the S-L equation has a direct bearing on the physics. For example, the eigenvalue Amay correspond to the orbital angular momentum of an electron in an atom (see the treatment of spherical harmonics in Chapter 12) or to the energy levels of a particle in a potential (see Example 14.5.2). In many cases, then, it is worthwhile to gain some knowledge of the behavior of an S-L system in the limit of large A-high angular momentum or high energy. Similarly, it is useful to understand the behavior of the solutions for large values of their arguments. We therefore devote this section to a discnssion of the behavior of solutions of an S-L system in the limit of large eigenvalues and large independent variable.

518

18. STURM-LIOUVILLE SYSTEMS: FORMALISM

18.3.1 Asymptotic Behavior for Large Eigenvalues We assume that the SoL operator has the form given in Equation (18.1). This can always be done for an arbitrary second-order linear DE by multiplying it by a proper function (to make it self-adjoint) followed by a Liouville substitution. So, consider an SoL systems of the following form: u" + [A - q(x)]u

PrOfer substitution

== u" + Q(x)u =

0

whereQ=A-q

(18.15)

with separated Bes of (18.2). Let us assume that Q(x) > 0 for all x E [a, b], that is, A> q(x). This is reasonable, since we are interested in very large A. The study of the system of (18.15) and (18.2) is simplified if we make the Priifer snbstitution: (18.16) where R(x, A)and (x, A)are A-dependent functions of x. This substitution transforms the SoL equation of (18.15) into a pair of equations (see Problem 18.3): d q' dx = .vA - q(x) - 4[A _ q(x)] sin2, dR = Rq' cos2. dx 4[A - q(x)]

(18.17)

The function R(x, A) is assumed to be positive because any negativity ofu can be transferred to the phase (x, A). Also, R carmot be zero at any point of [a, b], because both u and u' would vanish at that point, and, by Lemma 13.3.3, u(x) = O. Equation (18.17) is very useful in discussing the asymptotic behavior of solutions of SoL systems both when A --> 00 and when x --> 00. Before we discuss such asymptotics, we need to make a digression.

It is often useful to have a notation for the behavior of a function j (x, A) for large Aand all values of x. If the function remains bounded for all values of x as A --> 00, we write j(x, A) = 0(1). Intuitively, this means that as Agets larger and larger, the magnitude of the function j(x, A) remains of order 1. In other words, for no value of x is lim"....co j(x, A) infinite. If A"j(x, A) = 0(1), then we can write j(x, A) = O(1)/A". This means that as A tends to infinity, j(x, A) goes to zero as fast as I/A" does. Sometimes this is written as j(x, A) = OrA-"). Some properties of 0(1) are as follows:

1. If a is a finite real number, then 0(1) + a = 0(1). 2. 0(1)

+ 0(1) =

0(1), and 0(1)0(1) = 0(1).

3. For finite a and b,

J:

0(1) dx = 0(1).

4. If r and s are real numbers with r O(I)A'

.'

+ O(I)A' = o (1)A' .

::s s, then

18.3 OTHER PROPERTIES OF STURM--L10UVILLE SYSTEMS

519

5. If g(x) is any bounded function of x, then a Taylor series expansion yields

[A + g(x)]' = A'

[I + g~)

, { =A

r

g(x) r(r I+r-+ A 2

I) [g(X)]2

-A

O(I)} +-A3

= A' + rg(x)A,-1 + O(l)A,-2 = A' + O(l)A,-1 = 0(1)).'. Returning to Equation (18.17) and expanding its RHSs using property 5, we obtain dR

dx

=

0(1)

A

Taylor series expansion of q,(x, A) and R(x, A) about x = a then yields q,(x, A) = q,(a, A) + (x - a)Vi + R(x, A) = R(a, A)

for A -+

00.

0(1)

Vi'

0(1) + -A-

(18.18)

These resnlts are useful in determining the behavior of An for large

n. As an example, we use (18.2) and (18.16) to write _ al

=

Ih

u'(a) u(a)

=

R(a, A)Ql/4(a, A) cos[q,(a, A)] R(a, A)Q 1/4(a, A) sin[q,(a, A)l

=

Ql/2(a, A)cot[q,(a, A)],

where we have assumed that,81 f' O. If,81 = 0, we can take the ratio ,81/al, which is finite because at least one of the two constants must be different from zero. Let A = -al/,81 and write cot[q,(a, A)] = A/.JA q(a). Similarly, cot[q,(b, A)] = B/.JA - q(b), where B = -a2/.82. Let us concentrate on the nth eigenvalue and

write
-1

A "fAn - q(a)

A

1

q,(b, An) = coC --r.;'==T,[i' .JAn q(b)

For large An the argument of cot" 1 is small. Therefore, we can expand the RHS in a Taylor series about zero: 1

cot" e for s =

ee

ccf"

1

1f

1f

0(1)

(0)-£+···=--£+···=-+-2 2 A.

o (I)/A.. It follows that 1f

q,(a, An) =

0(1)

"2 + A.'

1f

q,(b, An) = -

2

+ nit +

0(1) 1'>'

vAn

(18.19)

520

18. STURM-LIOUVILLE SYSTEMS: FORMALISM

The term nit appears in (18.19) because, by Theorem 18.1.9, the nth eigenfunction has n zeros between a and b. Since u = R Q~ 1/4 sin tP, this means that sin tP must go through n zeros as x goes from a to b. Thus, at x = b the phase tP must be nn

largerthan at x

= a.

Substituting x = b in the first equation of (18.18), with A --> An, and using (18.19), we obtain

~ +n1l' + 2

0(1) =

.n;

~+ 2

+ (b -a)J):;; + 0(1),

0(1) .;r;;

.n;

or

(b - a)J):;; = nn

+

0(1)

(18.20)

1'>'

vAn

One consequence of this result is that, limn-->oo nA;I/2 = (b-a)I1l'. Thus,.;r;; = Csn, where limn-->oo Cn = 1l'1(b - a), and Equation (18.20) can be rewritten as

J):;; = ~ + 0(1) = ~ + 0(1). b -a

Cnn

b -a

n

(18.21)

This equation describes the asymptotic behavior of eigenvalues. The following theorem, stated without proof, describes the asymptotic behavior ofeigenfunctions.

18.3.1. Theorem. Let {un(x)} ~o be the normalized eigenfunctions ofthe regular S-L system given by Equations (18.15) and (18.2) with fhfh i' O. Then.for n --> asymptotic behavior of solutions of large order

00,

_~

unix) -

--cos

b-a

n1l'(x-a) b-a

0(1) + --. n

18.3.2. Example. Let us derive an asymptotic formula for the Legeudre polynomials Pn (x). We firstmake the Liouville substitution to transform the Legendre DE [(I-x 2) pI]' + n(n + I)Pn = 0 ioto d 2v dt 2

+ [An -

Here p(x) and

=

Pn(x(t))

where An = n(n + I).

Q(t)]v = 0,

1- x 2 and w(x)

=

I, so t

= v(t)[I-x2(t)]-1/4 =

=

r ds/~ = cos-I x, or x(t) = cost,

;(1) . smr

10 Equation (18.22)

(18.22)

(18.23)

16.3 OTHER PROPERTIES OF STURM--LIOUVILLE SYSTEMS

For large n we can neglect Q(t), make the approximation An ~ (n

v + en + ! )2v = 0, whose general solution is v(t) = A cos[(n + ~)t

521

+ !)2, and write

+ "J,

where A and a are arbitrary constants. Substituting this solution in (18.23) yields Pn(cos t) = A cos[(n + ~)t + "J/v'sint. To determine" we note that Pn(O) = 0 if n

is odd.Thus,if we let t = 1r/2, the cosine term vanishesfor oddn if a = -JC/4. Thus,the general asymptotic formula for Legendre polynomials is Pn(cost) =

I v" A cos [ (n + -)t

2

smt

-"] .

III

4

18.3.2 Asymptotic Behavior for Large x Liouville and Pnifer substitutions are useful in investigating the behavior of the solutions of S-L systems for large x as well. The general procedure is to transform the DE into the form of Eqnation (18.8) by the Liouville substitution; then make the Pnifer substitution of (18.16) to obtain two DEs in the form of (18.17). Solving Equation (18.17) when x -+ 00 determines the behavior of

(2

[

(

(2 . [

(

1)](2+ v 2x1/4] + 0(1) x5/2' 2

Jv(x)=y;;;;COS x- v+2: Yv (x) =

y;;;; sm

x -

-

2 1)]( v -1/4] 0(1) v + 2: 2 + 2x + x5/2 .

These two relations easily yield the asymptotic expressions for the Hankel functions:

HP)(x) = Jv(x) =

+ iYv(x)

{!; {[ (

1)](-+ v -1/4]} +--, 0(1) 2

-exp i x- v+](X

2

2

x 5/ 2

2x

H~2)(x) = Jv(x) - iYv(x)

=

(2 exp

Y;;;;

{-i [x _(v + ~)2 ::.2 + v

2 -

1/4]} + 0(1). x 5/2

2x

If the last term in the exponent-which vanishes as x -+ oo-is ignored, the asymptotic expression for HPJ (x) matches what was obtained in Chapter 15 using the method of steepest descent.

522

18. STURM-LIOUVILLE SYSTEMS: FORMALISM

18.4 Problems 18.1. Show that the Liouville substitutiou trausfonns regular S-L systems into regular S-L systems aud separated aud periodic BCs into separated aud periodic BCs, respectively. 18.2. Let UI (x) aud U2(X) be trausfonned, respectively into VI(t) aud V2(t) by the Liouville substitutiou. Show that the inner product ou [a, b] with weight functiou w(x) is trausfonned into the inner product ou [0, c] with unit weight, where c =

J:

../w/pdx.

18.3. Derive Equatiou (18.17) from (18.15) using Priifer substitution. 18.4. (a) Show that the Liouville substitution trausfonns the Bessel DE into 2

d v

-2

dt

+

(2

k -

v

2

-

t

2

1/4) _

v-D.

(b) Find the the equations obtained from the Priifer substitution, aud show that for large x these equations reduce to

, (

a) + ~ 0(1)

R'

= k I - 2k 2x2

R

=

0(1) ~'

t.

where a = v 2 (c) Integrate these equations from x to b > x aud take the limit as b --+

a (x) = 00 +kx+ 2kx

0(1)

00

to get

0(1) R(x) = Roo + - 2 - '

+7'

x

=

=

where 00 limb-->oo( (b) - kb) aud Roo limb-->oo R(b). (d) Substitute these aud the appropriate expression for Q-I/4 in Equation (18.16) aud show that 2

vex) =

1/4) +

v Roo ( Ii: cos kx - kxoo + -Jk 2kx

0(1)

-2-'

x

where kxoo ea n /2 - 00' (e) Choose Roo = ../2/:n: for all solutions of the Bessel DE, aud let aud for the Bessel functions Jv(x) aud the Neumann functions Yv (x), respectively, aud find the asymptotic behavior of these two functions.

IB.4 PROBLEMS

523

Additional Reading 1. Birkhoff, G. and Rota, G.-C. Ordinary Differential Equations, 3rded., Wiley, 1978. Has a good discussion of Sturm-Liouville differential equations and their asymptotic behavior.

2. Boccara, N. Functional Analysis, Academic Press, 1990. Discusses the Sturm-Liouville operators in the same spirit as presented in this chapter. 3. Hellwig, G. Differential Operators of Mathematical Physics, AddisonWesley, 1967. An oldie, but goodie! It gives a readable account of the Sturm-Liouville systems.

19

_

Sturm-Liouville Systems: Examples

Chapter 12 showed how the solution of many PDEs can be written as the product of the solutions of the separated ODEs. These DEs are nsually of Sturm-Liouville type. We saw this in the construction of spherical harmonics. In this chapter, consisting mainly of illustrative examples, we shall consider the use of other coordinate systems and construct solutions to DEs as infinite series expansions in terms of SoL eigenfunctions.

19.1 Expansions in Terms of Eigenfunctions Central to the expansion of solutions in terms of SoLeigenfunctions is the question of their completeness. This completeness was established for a regular SoLsystem in Theorem 18.1.9. We shall shortly state an analogous theorem (without proof) that establishes the completeness of the eigenfunctions of more general SoL systems. This theorem requires the following generalization of the separated and the periodic BCs: RIU

== anura) + a12u'(a) + anu(b) + aI4u'(b) =

0,

Rzu

sa aZlu(a)

+ «nu'ta) + aZ3u(b) + a24u'(b) =

0,

where

aij

(19.1)

are numbers such that the rank of the following matrix is 2:

TheseparatedBCscorrespondtothecaseforwhichan = aj, a12 = fh,az3 = az, and aZ4 = th, with all other aij zero. Similarly, the periodic BC is a special case for which an = -an = azz = -a24 = I, with all other aij zero. It is easy to

19.1 EXPANSIONS IN TERMS OF EIGENFUNCTIONS

525

verify that the rank of the matrix a is 2 for these two special cases. Let

ll= {u E eZ[a,bJIRju =0,

for j = 1,2}

(19.2)

bea subspace of.c~ (a, b), and-to assure the vanishing oftheRHS of the Lagrange identity-assume that the following equality holds:

P

(b) det

("'"'ZIU "'"'2212) = P (a) det ("'"'Z3t3

"'t4). "'24

(19.3)

We are now ready to consider the theorem (for a proof, see [Hell 67, Chapter 7]).

19.1.1. Theorem. The eigenfunctions (un(x)}~1 of an S-L system consisting of the S-L equation (pu')' + (Aw - q)u = 0 and the Bes of(19.1)form a complete basis of the subspace II of .c~(a, b) described in (19.2). The eigenvalues are real and countably infinite and each one has a multiplicity of at most 2. They can be ordered according to size Al ::: Az ::: .. " and their only limit point is +00. First note that Equation (19.3) contains both separated and periodic BCs as special cases (problem 19.1).10 the case of periodic BCs, we assume that pea) = p (b). Thus, all the eigenfunctions discussed so far are covered by Theorem 19.1.1. Second, the orthogonality ofeigenfunctions corresponding to different eigenvalues and the fact that there are infinitely many distinct eigenvalues assure the existence of infinitely many eigenfunctions. Third, the eigenfunctions form a basis of II and not thewhole.c~(a, b). Onlythosefunctionsu E .c~(a, b) thatsatisfytheBCin(19.1) are expandable in terms of u,,(x). Finally, the last statement of Theorem 19.1.1 is a repetition of part of Theorem 18.1.9 but is included because the conditiuns under which Theorem 19.1.1 holds are more general than those applying to Theorem 18.1.9. Part II discussed orthogonal functions in detail and showed how other functions can be expanded in terms of them. However, the procedure nsed in Part II was ad hoc from a logical standpoint. After all, the orthogonal polynomials were invented by nineteenth-century mathematical physicists who, in their struggle to solve the PDEs of physics using the separation of variables, carne across various ODEs of the second order, all of which were recognized later as S-L systems. From a logical standpoint, therefore, this chapter shonld precede Part II. But the order of the chapters was based on clarity and ease of presentation and the fact that the machinery of differential equations was a prerequisite for such a discussion. Theorem 19.1.1 is the important link between the algebraic and the analytic machinery of differential equation theory. This theorem puts at our disposal concrete mathematical functions that are calculable to any desired accuracy (on a computer, say) and can serve as basis functions for all the expansions described in Part II. The remainder of this chapter is devoted to solving some PDEs of mathematical physics using the separation of variables and Theorem 19.1.1.

526

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

t:P= f(x,y)

y

c b

x

Figure 19.1 A rectangular conducting box of which one face is held at the potential f (x, y) and the otherfaces are grounded.

19.2 Separation in Cartesian Coordinates Problems most suitable for Cartesian coordinates have boundaries with rectangular symmetry such as boxes or planes. rectangular conducling box

19.2.1. Example.

RECrANGULAR CONDUCTING BOX

Consider a rectangular conducting box with sides a, b, and c (see Figure 19.1). All faces are held at zero potential except the top face, whose potential is given by a function f(x, y). Let us find the potential at all points inside the box. The relevant PDEfor this situation is Laplace's equation, V 2 ep = O.Writing cJl(x, y, z) as a product of three functions, (x, y, z) = X(x)Y(y)Z(z), yields three ODEs (see

Problem 19.2): d 2Z

d 2X

- 2 +vZ=O,

- 2 +AX=O,

dz

dx

where A + f.l-

+v =

(19.4)

O.The vanishing of cI> at x = 0 and x = a means that

<1>(0, y, z) = X(O)Y(y)Z(z) = 0

V y, z

=}

X(O) = 0,

(a, y, z) = X(a)Y(y)Z(z) = 0

V y, z

=}

X(a) = O.

+ AX = 0, X(O) = 0 = X(a), whose Be is neither separated nor periodic, but satisfies (19.1) with au = Q(23 = 1 and all other aii zero. This S~L system has the eigenvalues and eigenfunctions

We thus obtain an S-L system, XIf

An = (n;)2

and

XIl(X)

= sin (n;x)

forn= 1,2, ....

Similarly, the second equation in (19.4) leads to

JLm=

(bmn)2

and

for m = 1,2, ....

On the other hand, the third equation in (19.4) does not lead to an S-L system because the Befor the top of the box does not fit (19.1). This is as expected because the "eigenvalue" v is already determined by A and JL. Nevertheless, we can find a solution for that equation. The substitution

2 _(,!!:)2 (mn)2 + b a

Ymn -

19.2 SEPARATION IN CARTESIAN COORDINATES

527

changes the Z equation to Z" - Y~n Z = 0, whose solution, consistent with Z (0) = 0, is Z(z) = Cmn sinh(Ymnz). We note that X(x) and Y(y) are functions satisfying RIX = 0 = R2X. Thns, by Theorem 19.1.1, they can be written as a linear combination of Xn(x) and YmCy): X(x) = I:~l An sin(mrxja) and Y(y) = I:;;,'=l B m sin(mrrbjy). Conseqnently, the most general solution can be expressed as 00

(x, y, z)

= X(x)Y(y)Z(z) = L

00

L

A mn sin

n=lm=l

c:

x) sin (mbrr y) sinh(YmnZ),

where A mn = AnBmCmn. To specify cI> completely, we must determine the arbitrary constants A mn . This is done f(x, y), yielding the identity by imposing the remaining BC, (x, y, c) 00

f(x, y) =

00

LL

n=lm=!

=

ff

n=lm=!

Amn sin

c:

=

x) sin

C C

rr

y) sinh(Ymn c)

rr

BmnsinC x) sin

y),

where Bmn == A mn sinh(Ymnc). This is a two-dimensional Fourier series (see Chapter 8) whose coefficients are given by 4 Bmn = ab

(nrr) (mrr) 10fa dx 1t' 0 dyf(x, y)sin --;;x sin bY .

III

Pierre Simon de Laplace (1749-1827) was a French mathematician and theoretical astronomer who was so famous in his own time that he was known as the Newton of France. His main interests throughout his life were celestial mechanics, the theory of probability, and personal advancement. At the age of 24 he was already deeply engaged in the detailed application of Newton's law of gravitation to the solar system as a whole, in which the planets and their satellites are not governed by the sun alone, but interact with one another in a bewildering variety of ways. Even Newton had been of the opinion that divine intervention would occasionally be needed to prevent this complex mechanism from degenerating into chaos. Laplace decided to seek reassurance elsewhere, and succeeded in proving that the ideal solar system of mathematics is a stable dynamical system that will endure unchanged for all time. This achievement was only one of the long series of triumphs recorded in his monumental treatise Mecamque Celeste (published in five volumes from 1799 to 1825), which summed up the work on gravitation of several generations of illustrious mathematicians. Unfortunately for his later reputation, he omitted all reference to the discoveries of his predecessors and contemporaries, and left it to be inferred that the ideas were entirely his own. Many anecdotes are associated with this work. One of the best known describes the occasion on which Napoleon tried to get a

528

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

rise out of Laplace by protestingthat he had written a huge bookon the systemof the world without once mentioning God as the author of the universe. Laplace is supposed to have replied, "Sire, I had no need of that hypothesis." The principal legacy of the Mecanique Celeste to later generations lay in Laplace's wholesale development of potential theory,

with its far-reaching implicationsfor a dozen different branches of physical science ranging from gravitation and fluidmechanics to electromagnetism and atomic physics. Even though he lifted the idea of the potential from Lagrange without acknowledgment, he exploited it so extensively that ever since his time the fundamental equation of potential theory has been known as Laplace's equation. After the French Revolution, Laplace's political talents and greed for position came to full flower. His compatriots speak ironically of his "suppleness" and "versatility" as a politician. What this really means is that each time there was a

change of regime (and there were many), Laplace smoothlyadapted himself by changing his principles-back and forth between fervent republicanism and fawning royalism-and each time he emerged with a better job and grander titles. He has been aptly compared with the apocryphal Vicar of Bray in English literature, who was twice a Catholic and twice a Protestant. The Vicar is said to have replied as follows to the charge of being a turncoat: "Not so, neither, for if I changed my religion, I am sure I kept true to my principle, which is to live and die the Vicar of Bray." To balance his faults, Laplace was always generous in giving assistance and encouragement to younger scientists. From time to time he helped forward in their careers such men as the chemist GaywLussac. the traveler and naturalist Humboldt, the physicist Poisson, andappropriately-the young Cauchy, who was destined to become one of the chief architects of nineteenth century mathematics.

Laplace's equation describes not only electrostatics, but also heat transfer. When the transfer (diffusion) of heat takes place with the temperature being independent of time, the process is known as steady-state heat transfer. The diffusion equation, aT fat = a2 V 2 T , becomes Laplace's equation, V 2T = 0, and the technique of the preceding example can be used. It is easy to see that the diffusion equation allows us to perform any linear transformation on T, such as T -> «T + P, and still satisfy that equation. This implies that T can be measured in any scale such as Kelvin, Celsius, and Fahrenheit. steady-state haat-conducting plate haat-conducting plata: steady state

19.2.2. Example.

STEADY-STATE HEAT-CONDUCTING PLATE

Let us consider a rectangular heat-conducting plate with sides of lengths a and b. Three of the sides are held at T = 0, and the fourth side has a temperaturevariation T = f (x) (see Figure 19.2). The flat faces are insulated. so they cannot lose heat to the surroundings. Assuming a steady-state heat transfer, let us calculate the variation of T over the plate. The problem is two-dimensional. The separation of variables leads to

where A + f.' = O. The X equationandthe BCs T(O, y) = I'(a, y) =

(19.5)

oform an S-L systemwhoseeigenvalues

and eigenfunctions are An = (mr/ap and X n (x) = sin(mrx/a) forn = 1,2, .... Thus.accordingtoTheorem19.1.1,ageneralX(x) canbewrittenasX(x) L:~t An sin(nrrxja).

=

19.2 SEPARATION IN CARTESIAN COOROINATES

529

T=f(x)

o· ---x

o· f-oE---- a Figure 19.2

301

------;~

A heat-conducting rectangular plate.

The Y equation, on theother hand, does not form anS-L systemdue to thefactthat its "eigenvalue" is predetermined by thethird equation in (19.5). Nevertheless, we cansolvethe equation Y" -(mr/apr = Otoobtainthegeneral solutionY(y) = Aen:Jry/a + Be-mry/a . Since T'[x, 0) = 0 'I x, we must have Y(O) = O. This implies thatA+B = 0, which.tnturn, reduces the solution to Y = A sinh(mryja). Thus, the most general solution, consistent with the three Bes T(O, y) = T(a, y) = T(x, 0) = 0, is 00

ri», y) = X(x)Y(y) =

L s; sin err x) sinh err y). n=l

a

a

The fourth Be gives a Fourier series, f(x)

~ [ s; smh . (nrr)]. = L--;;b sm (nrr) --;;x 11=1

. sa ~ L- en sin

(nrr --;;x) ,

n=l

whose coefficientscan be determined from nrr ) c, = s; sinh ( --;;b =;;2 10fa sin (nrr) --;;x f(x) dx. In particular, if thefourth side is held atthe constant temperature

2T. (.!'-) [1- (_1)n] = { c; = --.!l. a

nH

4 TO

nrr

0

To. then

if n is odd, if n is even,

and we obtain

f

T(x, y) = 4To 1 sin[(2k n k~O 2k + I

+ 1)rrx/alsinh[(2k + I)rry/al. sinh[(2k

+ I)rrb/a]

(19.6)

If thetemperature variation of thefourth sideis of theform f (x) = To sin(n x/a). then

en = -2To Ina sin (nrrx) - - sin (rrx) dx = -2To (a) - 8n 1 = To8 n 1 a 0 aa a 2' ,

530

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

and Bn = T(x

en / sinh(mrb/a) = )

,y

= To 0

[To/ sinh(mrb/a)]8n," and we have

sin(rrx/a) sinh(rry/a). sinh(rrb/a)

(19.7)

Only one term of the series survives in this case because the variation on the fourth side happens to he one of theharmonics of theexpansion. ' Note that the temperature variations given by (19,6) and (19.7) are independent of the material of the plate because we are dealing with a steady state. The conductivity of the material is a factor only in the process of heat transfer, leading to the steady state.

Onceequilibrium hasbeenreached, the distribution of temperature will be the sameforall materials. II The preceding two examples concerned themselves with static sitoations. The remaining examples of this section are drawn from the (time-dependent) diffusion equation, the Schrodinger equation, and the wave equation. conduction 01 heal in a rectangular plate

19.2.3. Example. CONDUCTION OF HEAT IN A RECTANGULAR PLATE Consider a rectangularbeat-conducting plate with sides oflength a and b all held at T = O. Assume that at time t = 0 the temperature has a distribution function I(x, y). Let us find the variation of temperature forall points (x, y) at all times t > O. The diffusion equation for this problem is aT at

= k 2V 2T = k 2 (a

2;

ax

+

2;). a ay

A separation of variables, T(x, y, t) = X(x)Y(y)g(t), leads to three DEs:

d2 y - 2 +iLY=O, dy

The BCs T(O, y, t) = T(a, y, r) = T(x, 0, t) = T(x, b, r) = 0, together with the three ODEs,give rise to two S-L systems.The solutions to bothof these are easily found: _ (mr)2 Ana _ (mrr)2 iLm- b

= sin (n; x)

and

Xn(X)

and

Ym(y) = siu

err

y)

for n = 1,2, ... , for m

= 1,2, ....

These give rise to thegeneral solutions 00

X(x)

= ~::>n sin ex), n=I

With Ymn == k 2Jr2(n 2/a2 + m 2/h2), the solutionto the g equation can be expressed as g(t) = Cmne-Ymnt. Putting everything together, we obtain T(x, y, t)

00 =E

L Amne-Ymnt sin (nrr) --;;x sin (mrr) bY , 00

n=lm=l

19.2 SEPARATION IN CARTESIAN COOROINATES

531

where Amn = AnBmCmn is an arbitrary constant. To determine it, we impose the initial condition T(x, y, 0) = f(x, y). This yields 00

f(x,y) =

00

L L A mn sin n=lm=l

C:

x ) sin

which determines the coefficients Amn : A mn

quantum particle in a box

= a: fa" dx

t

C"

C:

dyf(x, y) sin

y),

x) sin

C"

y).

..

19.2.4. Example. QUANTUM PARTICLEIN A BOX Thebehavior of anatomic particle of mass Jl- confined in a rectangular box withsides a, b, andc (aninfinite three-dimensional potential well) is governed bythe Schrodinger equation for a freeparticle,

2

2

2

2

ai

az 2

+ a 1/1 + a 1/1 ) ,

ina1/l = _ n (a 1/1 at 2/L ax 2

and the Be that 1/I(x, y, Z, t) vanishes at all sides of the boxfar all time. A separation of vanabies 1/1 (x, y, z, t) = X(x)Y(y)Z(z)T(t) yields the ODEs d 2X - 2 +).X=O, dx dT

d 2Z

d 2y - 2 +uY=O,

- 2 +vX=O,

dy

-+iwT=O, dt

where

ni

n ea -().+u+v).

dz

2/L

The spatial equations, together with the Bes

1/1(0, y, Z, t) = 1/I(a, y, Z, t) = 0 1/1 (x, 0, z, t) = 1/I(x, b, z, t) = 0 1/1 (x, y, 0, t) = 1/I(x, y, c, t) = 0

=}

X(O) = 0 = X(a),

=}

YeO) = 0 = Y(b),

=}

Z(O) = 0 = Z(c),

lead to three S-L systems, whose solutions areeasily found:

(n; x),

n")2' (a m" )2 ' m = (b

An =

for n = 1,2, ... ,

Ym(y) = sin (m " y), b

U

for m = 1,2, ...•

ZI(Z) = sin

VI

Xn(X) = sin

ez),

=

C:)2

for I

= 1,2, ....

The time equation, on the other hand, bas a solutionof the form where

Wimn = 2:

[C)2 + C")2 +

en·

The solution of the Schrodinger equation that is consistent withthe Bes is therefore 1fr(x, y, Z, r)

=

f:

1,m,n=1

Almne-iWlmnt sin

(n; x) sin (mb Y) sin 7r

c:

z).

532

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

The constants Almn are determined by the initial shape, 1/1 (x, y, Z, 0) of the wave function. The energy of the particle is

/i2,,2 E=nwl mn = - 2/L

2 (n-2+2+2 m2 12 ) .

abc

Each set of three positive integers (n, m,l) represents a state of the particle. For a cube, a = b = c == L, and the energy of the particle is

/i2,,2 /i2,,2 2 2 2 E = - - 2 (n 2 + m + z2) = ------Z-/3 (n + m + z2) 2/LL 2/LY

(19.8)

where V = L 3 is the volume of the box. The ground stale is (1,1,1), has energy E = 3lt2rr 2 /2f.l- V 2/3, and isnondegenerate(only one state correspondsto this energy). However, thehigher-level statesaredegenerate. Forinstance,the threedistinctstates(I, 1,2), (I, 2, I), and (2,1,1) all correspond to the same energy, E = 6/i2,,2/2/Ly 2/3. The degeneracy increases rapidly with larger values ofn, m, and I. Equation(19.8) can be written as

where

R

2

2/LEy 2/3 =

h,2

1r2

This looks like the equation of a sphere in the nml-space.1f R is large, the number of states contained within the sphere of radius R (the number of states with energy less than or equal to E) is simply the volume of the first octant! of the sphere. If N is the number of such states, we have

density of states

Thus the density of states (the number of states per unit volume) is _ N _::. n- y - 6

Fermi energy

(~)3/2 3/2 /i2,,2

E.

(19.9)

This is an important formula in solid-state physics, because the energy E is (with minor modifications required by spin) the Fermi energy. If the Fermi energy is denoted by E f' Equation (19.9) gives E f = an 2j3 where a is some constant. l1li

In the preceding examples the time variation is given by a first derivative. Thns, as far as time is concerned, we have a FODE. It follows that the initial specification of the physical quantity of interest (temperature T or Schrodinger wave functiou 1/1) is sufficient to determine the solution uniquely. A second kind of time-dependent PDE occurring in physics is the wave equation, which contains time derivatives of the second order. Thus, there are two arbitrary parameters in the general solution. To determine these, we expect two initial conditions. For example, if the wave is standing, as in a rope clamped at both IThis is because n, m, and 1 are all positive.

19.2 SEPARATION IN CARTESIAN COOROINATES

533

ends, the boundary conditions are not sufficient to determine the wave function uniquely. One also needs to specify the initial (transverse) velocity ofeach point of the rope. For traveling waves, specification of the wave shape and velocity shape is not as important as the mode of propagation. For instance, in the theory of wave guides, after the time variation is separated, a particular time variation, such as e+ iwt , and a particular direction for the propagation of the wave, say the z-axis, are chosen. Thus, if u denotes a component of the electric or the magnetic field, we can write u(x, y, Z, z) = 1fr(x, y)e'(WI±k'l, where k is the wave number. The

waveequation thenreduces to

Introducing y2 = w2/c2 - k 2 and the transverse gradient V, = (a/ax, a/ay) and writing the above equation in terms of the full vectors, we obtain

{E}

(V,2 + Y 2 ) B = 0,

where

y)} e,(wIHz) {BE}= {E(x, B(x, y)

(19.10)

These are the basic equations used in the study of electromagnetic wave guides

andresonant cavities. guided waves

Maxwell's equations in conjunction with Equation (19.10) gives the transverse components (components perpendicular to the propagation direction) E , and B, in terms of the longitudinal components E z and B z (see [Lorr 88, Chapter 33]):

(aEz ) - ,-e .W. x (V,B ), z z az c 2 .W. , (aB yB,=V - z ) +,-e,x(V,E z)' y 2 E, = V, -

az

c

(19.11)

Three types of guided waves are usually studied. 1. Transverse magnetic (TM) waves have B, = 0 everywhere. The BC on E demands that E z vanish at the conducting walls of the guide. 2. Transverse electric (TE) waves have E, = 0 everywhere. The BC on B requires that the normal directional derivative

vanish at the walls. 3. Transverse electromagnetic (TEM) waves have B, = 0 = E z . For anontrivial solution, Equation (19.11) demands that y2 = O. This form resembles a free wave with no boundaries.

534

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

We will discuss the TM mode briefly (see any book on electromagnetic theory for further details). The basic equations in this mode are

('v; + YZ)E z = Z

Y E, = V, rectangular wave guides

B, =0,

0,

(8E,) az '

yZB, = i :':e, c

x (V,E,).

(19.12)

19.2.5. Example. REcrANGULAR WAVE GUIDES For a wave guide with a rectangular cross section of sides a and b in the x and the y directions, respectively, we have aZE,

aZE

ax

ay

Z

-z- + -z-z + Y E, = O. A separationof variables, E,(x, y)

dZX

= X(x)Y(y), leads to two SoL systems,

= 0 = X(a),

- Z +AX=O,

X(O)

- Z +/LY=O,

Y(O) = 0 = Y(b),

dx dZY dy

where y2 = A + u: Theseequations havethe solutions Xn(X)

= sin

Ym(y) =

(n; x) ,

sin (";,1!' y) ,

_ (n1!')Z

An -

u-m

for n = 1,2, ... ,

a

=

(";,1!')Z

for m

=

1,2, ....

The wave number is given by

which has to be realif the waveis to propagate (an imaginary k leadsto exponential decay or growthalong the z-axis). Thus, there is a cutoff frequency, {orm,n

2: 1,

below which the wave cannot propagate through the wave guide. It follows that for a TM wave the lowest frequency that can propagate along a rectangular wave guide is WI! 1!'cJa Z + bZj(ab). The most general solutionfor E z is therefore 00

Ez

=

L m,n=l

Amnsin(n; X) sin

(mb

1r

y)ei(wt±kmnZ).

The constants Amn are arbitrary andcan bedetermined from the initial shapeof the wave, butthatis not commonly done. Once E z is found, the othercomponents can be calculated using Equation (19.12). II

19.3 SEPARATION IN CYLINDRICAL COORDINATES

535

«P=V(p,
T h

y x Figure 19.3 A conducting cylindricalcan whose top has a potential given by V(p, \,), with the rest of the surfacegrounded.

19.3 Separation in Cylindrical Coordinates When the geometry of the boundaries is cylindrical, the appropriate coordinate system is the cylindrical one. This usually leads to Bessel functions "of some kind." Before working specific examples of cylindrical geometry, let us consider a question that has more general implications. We saw in the previous section that separation of variables leads to ODEs in which certain constants (eigenvalues) appear. Different choices of signs for these constants Canlead to different functional forms of the general solution. For example, an equation such as d 2 x / dt 2 - kx = 0 can have exponential solutions if k > 0 or trigonometric solutions if k < O. One cannot a priori assign a specific sign to k. Thus, the general form of the solution

is indeterminate. However, once theboundary conditions areimposed, theunique solutions will emerge regardless of the initial functional form of the solutions (see [Hass 99] for a thorough discussion of this point). conducting cylindrical can

19.3.1. Example. CONDUCTING CYUNDRICAL CAN Consider a cylindrical conducting can of radius a and height h (see Figure 19.3). The

potential varies atthetopface as V (p,
2 d (dR) p - + (2 k p - -m ) R = 0, dp dp p

-

536

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

wherein anticipation of the correct Bes, we have written the constants as k 2 and _m 2 with m aninteger. Thefirst of these is theBesselequation, whosegeneral solution canbe written as R(p) = AJm(kp) + BYm(kp). The second DE, when the extra condition of periodicity is imposed on thepotential, has thegeneral solution

S(
Wenote that none of the three ODEsleadto an S-L system of Theorem 19.1.1 because the Bes associatedwiththemdo not satisfy (19.1). However, we can still solve the problem by

imposing thegiven Bes. The factthatthepotential mustbe finiteeverywhere inside thecan (including at p = 0) forces B to vanish because the Neumann function Ym (kp) is not defined at p = O. On the other hand, we want to vanish at p = a. This gives Jm(ka) = 0, which demands that ka be a root of the Bessel functionof order m. Denoting by Xmn the nth zero of the Bessel function of order m, we haveka = Xmn, or k = xmn/a for n = 1,2, .... Sintilarly, the vanishing of at Z = 0 implies that E=-F

Fourier-Bessel series

. (xmnz) Z(z) = E sinh -a- .

and

Wecan now multiply R, S, and Z and sumoverall possible values ofm and n, keeping inmind that negative values of m giveterms that arelinearly dependent onthecorresponding positive values. Theresult is theso-calledFourier-Bessel series: (p,
LL 00

00

m~On~l

mn ) . mn) . Jm (x . ---;;-P sinh (x (A mn cosm


---;;-z

where Amn and Bmn are constants to be determined by theremaining Be. To find these constants we use the orthogonality of the trigonometric andBessel functions. For z = h Equation (19.13) redoces to

yep, lp) =

f f: s;

n (x: p) sinh (x:n h) (A mn cosmlp + Bmn sinmlp),

m=On=l

from which we obtain A mn = Bmn =

2 2

2

22

2

x a Jm+ 1(xmn) sinh(xmnh/a) x a Jm+ 1(xmn) sinh(xmnh/a)

r:

Jo

d


{21r dlp

Jo

Jo

e

mn p) cosmtp, a

(a dPPV(p,lp)Jm(Xmnp)sinmlp,

Jo

a

where we haveusedthefollowing result derived in Problem 14.39: (19.14)

19.3 SEPARATION IN CYLINDRICAL COORDINATES

537

Forthe specialbutimportant case.of azimuthal symmetry, for which V is independent of sp, we obtain A mn =

m

40 ' 0 a 2Jf(xon) sinh(xOnhja)

In" 0

(x on )

dppV(p)Jo - p , a

.

Bmn =0.

The reason we obtained discrete values for k was the demand that vanish at

p = a. If we let a -+ 00, then k will be a continuous variable, and instead of a sum over k, we will obtain an integral. This is completely analogous to the transition froma Fourier series to a Fourier transform, but we will not pursue it further. circular heat-conducting plate

19.3.2. Example. CIRCULAR HEAT-CONDUCTING PLATE Consider a circular heat-conducting plate of radius a whose temperature at time t = 0 has a disIribution function f (p, 0 whenthe edge is keptat T = O. Thisis a two-dimensional problem involvingthe heatequation,

[.!.i.

aT = k 2V 2T = k 2 (/T) at p ap Bp

2T].

+ 2. a

p2 B
A separationof variables, T(p,
2S d - 2 +tLS=O,

2

-d =k Ag,

dsp

t

a d

2R p2

t dR

+ Pdp

-

(tL

p2

)

+A

R=

O.

To obtain exponential decay rather than growth for the temperature, we demand that A == b2 < O. To ensureperiodicity (see the discussion at the beginningof this section), we must have f.L = m 2, where m is an integer.'To have finite T at p = 0, no Neumann function is to be present. Thisleadsto the following solutions: get)

= Ae-k'b",

S(
= Bcosm


R(p)

= DJm(bp).

If the temperature is to be zero at p = a, we musthave Jm(ba) follows thatthe general solution canbe written as 00

T(p, sp, t)

=

xmn/a. It

00

L L e-k2(Xmn/a)2t Jm(x: n p){Amn cosm


A mn and Bmn canbe determined as in thepreceding example.

cylindrical wave guide

= 0, or b =

.

19.3.3. Example. CYLINDRICAL WAVEGUIDE Fora TM wavepropagating along the z-axis in a hollow circ~ar conductor, we have [see Eqnation(19.12)] 2E 1 B (BEZ ) t B z 2 - - p - + - - - + y Ez=O. p Bp Bp p2 B
538

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

The separation E z d d

2R p2

= R(p)S(rp) yields S(rp) = A cosmrp + H sinmrp and

I dR

+ Pdp +

(

2) m y2 - p2 R

= O.

The solution to thisequation, which is regular at p = 0 andvanishes at p = a, is R(p) = CJm(x mn p)

and

a

Xmn

y=-. a

Recalling thedefinition of y, we obtain

This gives the cut-off frequency lVmll

= cXmn/a.

The solution forthe azimuthally symmetric case (m = 0) is Ez(p, v, t)

f:

=

AnJO

(X~n p) ei(wt±knz)

and

Hz =0,

n=l

11II

There are many variations on the theme of Bessel functions. We have encounlered three kinds of Bessel functions, as well as modified Bessel functions. Another variation encountered in applications leads 10 whal are known as Kelvinfunctions, introduced in the following example.

current dlstrlbutlon in a circular wire

19.3.4. Example.

CURRENT DISTRlBUTION IN A CIRCULAR WIRE

Consider the flow of charges in an infinitely long wire with a circular cross section of radius Q. We are interested in calculating the variation of the current density in the wire as a function of time and location. Therelevant equation can be obtained by starting with Maxwell's equations for negligible charge density (V· E = 0), Ohm's law G = ,,-E), the assumption of high electrical conductivity (I,,-EI » jaE/atl), and the usual procedure of obtaining the waveequation fromMaxwell's equations. Theresult is 2. V

4",,,- aj

J-7at =0.

Moreover, we make the simplifying assumptions that the wire is along the z-axis and that there is no turbulence, so j is also along the z direction. We further assume that j is independent of


skindeplh Kelvin equation

I d .

+ _-.-1. + <2j pdp

= 0,

(19.15)

where <2 = i4"',,-w/c2 == i2/~2 and ~ = c/-'/2",,,-w is called the skin depth. The Kelvinequation is usually givenas d 2w I dw 2 - -2+ - - - i k w=O. dx x dx

(19.16)

.t "-.

19.3 SEPARATION IN CYLINDRICAL COORDINATES

Kelvin function

539

If we substitute x = ,Ji t/ k, it becomes in + tiJ / t + w = 0 whichis a Bessel equation of order zero. If the solution is to be regular atx = 0, thenthe only choice is wet) = Jo(t) = Jo(e- i -n j 4 kx). This is the Kelvin function for Equation (19.16). It is usually written as

Jo(e- hr/4kx) es ber(kx) + i bei(kx)

where berand bei stand for "Besselreal" and "Besselimaginary," respectively. If we substitute z = e- i1C j 4kx in the expansion for Io(z) and separate the real andthe imaginary parts of the expansion, we obtain ber(x) = 1-

. beifx)

=

(xI2)4 - +(xI2)8 - - ... (2!)2 (4!)2

(xI2)2 (xI2)6 (1!)2 - (3!)2

+

(xI2)10 (5!)2 _ ....

Equation (19.15) is the complex conjugate of(l9.16) withk 2 = 2/13 2 . Thus, its solution is

j(p) = AJo(ein/4kp) =-A {ber

(~ p) -

ibei

(~ p)}.

Wecancompare thevalue of thecurrent density at p with its value atthe surface p = a:

T P)]1/2 (T~ P) + bei2 (~ Ij(a) I [ ber2 (~) (~13 ) - a + b. el 2 - a 13 j(p)

ber

2

=

For low frequencies, 13 is large, which implies that P113 is small; thus, ber( ~p113) '" I and bei(~pllJ) '" 0, and Ij(p)lj(a)1 '" I; i.e., the current density is almost uniform. For higher frequencies the ratio of the current densities starts at a valueless than 1 at p = 0 and increases to 1 at p = a. The starting value depends on the frequency. For very large frequencies the starting value is almost zero (see [Mati 80, pp 150-156]). II

quanlum particle ina cylindrical can

19.3.5. Example. QUANTUM PARTICLE IN A CYUNDRICAL CAN Let us consider a quantum particle in a cylindrical can. For an atomicparticle of mass f.L confined in a cylindrical can of length L andradius a, the relevant Schriidinger equation is . at I

at

Ii = - 2{Jo

[IP a (

at) ap p ap

I +

2t a p2 arp2

+

a2t] az 2 .

Let us solve this equation subject to the Be that 1/!(p, qJ, Z, t) vanishes at the sides of the can. A separation of variables, t(p, rp, z, t) = R(p)S(rp)Z(z)T(t), yields dT -=-;OJT dt 2 d R dp2

d 2Z -+AZ=O, dz 2

+

d 2S

-+m2 S = O,

drp2 2 .!:. dR + (2{Jo OJ _ A _ m ) R = O. pdp Ii p2

(19.17)

540

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

The Z equation, along withits periodic Bes, constitutes an S-L systemwithsolutions

Z(z) = sin (k; z)

fork=I.2.....

If we let 2/Lw/1i- (k:n:/ L)2 '" b2• then the last eqnationin (19.17)becomes

2R 2 d +.!:.dR +(b2_ m ) R = O. d p2 pdp . p2

and the solutionthat is well-behaved at p = 0 is Jm(bp). Since R(a) = O. we obtain the quantization condition b = xmn/a for n = 1,2, .... Thus, theenergy eigenvalues are

and thegeneral solution canbe written as

~ e -iwkm. t Jm (xmn). . '!J(P.
a

III

L

19.4 Separation in Spherical Coordinates Recall that most PDEs encountered in physical applications can be separated. in spherical coordinates. into

L2y(O.
+~dR +[f(r)_I(I+I)]R=O. r dr

(19.18)

r2

We discussed the first of these two equations in great detail in Chapter 12. In particular. we constructed Yim(0.
d 2Sm

--2 d
-d [ (l-x 2 )dPlm] -dx dx

+m

2

Sm=O.

2

m ] + [ 1(1+1)--l-x 2

Plm=O.

where x = cos O. These are both S-L systems satisfying the conditions ofTheorem 19.1.1. Thus. the Sm are orthogonal among themselves and form a complete set for ,(,2(0. 2IT). Similarly. for any fixed m. the Plm(X) form a complete orthogonal set for ,(,2( -1. +1) (actually for the subset of ,(,2( -1. +1) that satisfies the same

19.4 SEPARATION IN SPHERICAL COORDINATES

541

BC as the Pi« do atx = ±I). Thus, the products Ylm(X,
2dR

1(1 + I)

dr

r dr

r2

-+ --2

R=O.

Multiplying by r 2 , substituting r = e', and using the chain rule and the fact that dt [dr = I/r leads to the following SOLDE with constant coefficients:

-d

2R

dt 2

dR + -·-1(1 + I)R = dt

O.

This has a characteristic polynomial peA) = A2 + A -1(1 + I) with roots Al = I and A2 = -(I + I). Thus, a general solution is of the fonu

+ Be A"

R(t) = Ae AI '

= A(e')l

+ B(e,)-I-I,

or, in tenus of r, R(r) = Ar1 + Br-I- I. Thus, the most general solution of Laplace's equation is 1

00

(r, e,
LL

(Alm r l

+ Blmr-1-I)Ylm(e,
I=Om=-l

For regions containing the origin, the finiteness of implies that Blm = O. Denoting the potential in such regions by in, we obtain 00

in(r,

e.
1

L L

AlmrIYlm(e,
I=Om=-1

Similarly, for regious including r = 00

out(r, e,
00,

we have

I

L L

Blmr-I-IYlm(e,
1=0 m=-l

To determine Aim and Blm, we need to invoke appropriate BCs. In particular, forinside a sphere of radius a on which the potential is given by Vee,
vee,
= in(a, e.
1

L

l=Om=-1

Alma1Ylm(e,
542

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

Y~ (0,

Multiplying by Akj

= a-k

II

91) and integrating over dQ = sin 0 dO d91, we obtain

dQV(O,

91}Y~(O, 91}

=} Aim

= a-I

II

dQV(O, 91}YI;"(O, 91}·

Similarly, for potential ontside the sphere, Rim

= a l+ 1

II

dQV(O, 91}YI;"(O, 91}·

In particular, if V is independent of 91, only the components for which m = 0

are nonzero, and we have 2rrI Am = a

1"

2rrI sinOV(O}Yio(O}dO = -

a

0

Fll-I + -1 1" smOV(O}PI(cosO} . -4 ae. tc 0

which yields

where Al = _2_ {" sinOV(O}PI(COSO} so. 21 + 1

10

Similarly,

Helmholtz equation

The next simplest case after Laplace's equation is that for which f (r) is a constant. The diffusion equation, the wave equation, and the Schriidinger equation for a free particle give rise to such a case once time is separated from the rest ofthe variables. The Helmholtz equationis (19.19) and its radial part is

2R d dr 2 spherical Bessel functions

+~dR r dr

+[k2_1(l+I}]R=0. r2

(19.20)

(This equation was discussed in Problems 14.26 and 14.35.) The solutions are spherical Bessel functions, generically denoted by the corresponding lower case letter as ZI(X} and given by (19.21)

543

19.4 SEPARATION IN SPHERICAL COOROINATES

where Zv(x) is a solution of the Bessel eqnation of order v. A general solution of (19.20) can therefore be written as

If the origin is included in the region of interest, then we must set B = O. For such a case, the solution to the Helmholtz equation is Z

00

LL

'h(r, IJ, 'P) =

(19.22)

Azmjz(kr)Yzm(lJ, 'P).

I=Om=-l

The subscript k indicates that'" is a solution of the Helmholtz equation with k2 as its constant.

particle Ina hard sphere

19.4.1. Example.

PARTICLE IN A HARD SPHERE

The time-independent Schrodinger equation for a particle in a sphere of radius a is

,,}

- 2jLv 2t = Et with the Be t(o, e, IfJ) = O. Here E is the energy of the particle

and jL is its mass. We rewrite the Schrodinger equation as V2t k2 = 2f.LE /;,,2,we can immediately writetheradial solution

+ 2jLEjn2

= O. With

Rz(r) = Ajz(kr) = Ajz(,j2jLErjn).

The vanishing of tat 0 impliesthat jZ(.,f2jLE ojn) = 0, or .,f2jLE 0 11,

for n

- = Xin

=

1,2, "',

where KIn is the nth zeroof hex), whichis the sameas the zeroof Jl+l/2(X). Thus, the energy is quantized as El

n

=

n2x~

for I = 0, 1, ... , n = 1,2, ....

2jLo

The general solution to the Schrodlnger equation is I

LL L 00

t(r,e,IfJ) =

00

n=11=Om=-1

r)

Anzmjz(Xznti YZm(e,'P).

III

A particularly useful consequence of Equation (19.22) is the expansion of a plane wave in terms of spherical Bessel functions. It is easily verified that if k is a vector, with k . k = k 2 , then e ik.r is a solution of the Helmholtz equation. Thus, e ik.r can be expanded as in Equation (19.22). Assuming that k is along the z-axis, we get k· r = krcoslJ, which is independent of 'P. Only the terms of Equation (19.22) for which m = 0 will survive in such a case, and we may write eik"osO = L~o A/j/(kr)Pz(cos IJ). To find Az, let u = cos IJ, mnltiply both sides by Pn(u),andintegratefrom-I to I:

1 . 1

Pn(u)e'krUdu =

-1

2 L Azjz(kr) 11 Pn(u)Pz(u) du = Anjn(kr)--. 2n + I 00

/=0

-1

544

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

Thus

11

. 2n + Anln(kr) = 2 = 2n

+1 2

1

Pn(u)e ikru du

-1

f

m=O

1

(19.23)

1

(ikr)m m!

Pn(u)umdu.

-1

This equality holds for all values of kr. In particular, both sides should give the same result in the limit of small kr. From the definition of jn (kr) and the expansion of I n(kr), we obtain

1

.Jii (kr)n

jn(kr) expansion ofeik.r in spherical harmonics

~ 2

2

['(n

+ 3/2)

On the other hand, the first nonvanishing term of the RHS of Equation (19.23) occurs when m = n. Eqnating these terms on both sides, we get 21..+1 n!

.Jii (kr)n An 2

2

(2n

+ 1)!.Jii

= 2n

+ 1 ;n(kr)n 2n+1(nl)2 2

n!

(2n

+ I)! '

(19.24)

where we have used I' ( n

~) _ (2n + 1) 1.Jii + 2 - 21..+1n!

and

1

1 -1

2n+1(n l)2 Pn(U)U du = (2n + 1)1' n

Equation (19.24) yields An = ;n(2n + 1). With An thns calculated, we can now write 00

elkroo'B = L(21

+ l);l j l (kr) PI(cos 6).

(19.25)

I~O

For an arbitrary direction of k, k . r = kr cos y, where y is the angle between k and r. Thus, we may write elk.• = I:Z:;o(21 + 1);1jl (kr) PI(cos V), and using the addition theorem for spherical harmonics, we finally obtain 00

elk -r = 4n

1

LL

;1h (kr)Y/;n (6', ,P')Ylm(6,
(19.26)

1=0 m=-l

where 6' and ip' are the spherical angles of k and 6 and


19.5 PROBLEMS

545

T=O

x

T=O Figure 19.4

A semi-infinite heat-conducting plate.

19.5 Problems 19.1. Show that separated and periodic Bes are special cases of the equality in Equation (19.3). 19.2. Derive Equation (19.4). 19.3. A semi-infinite heat-conducting plate of width b is extended along the positive x-axis with one comer at (0, 0) and the other at (0, b). The side of width b is held at temperature To, and the two long sides are held at T = 0 (see Figure 19.4). The two flat faces are insulated. Find the temperature variation ofthe plate, assuming equilibrium. Repeat the problem with the temperature of the short side held at each of the following:

Ca)T={O

To

(c) To cos

~f 0< y < biZ, If biZ < y < b.

Gy) ,

o :'0 y :'0 b.

(b)

;y,

(d) To sin

0:'0 y:'O b.

(~y)

,

0:'0

y:'0 b.

19.4. Find a general solution for the electromagnetic wave propagation in a resonant cavity, a rectangular box of sides 0 :'0 x :'0 a, 0 :'0 Y :'0 b, and 0 :'0 z :'0 d with perfectly conducting walls. Discuss the modes the cavity can accommodate. 19.5. The lateral faces of a cube are grounded, and its top and bottom faces are held at potentials 11 (x, y) and fz(x, y), respectively. (a) Find a general expression for the potential inside the cube. (b) Find the potential if the top is held at Vo volts and the bottom at - Vo volts. 19.6. Find the potential inside a semi-infinite cylindrical conductor, closed at the nearby end, whose cross section is a square with sides oflength a. All sides are grounded except the square side, which is held at the constant potential Vo.

(

546

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

19.7. Find the temperature distribution of a rectangular plate (see Figure 19.2) with sides of lengths a and b if three sides are held at T = 0 and the fourth side has a temperature variation given by

To a

(a) -x,

To a

0::; x < a.

I -"2 a I'

o::: x

To x (e) -;

0:" x :" a.

(b) 2"x(x - a), (d) T = 0,

< a.

0 :" x :" a.

19.8. Consider a thin heat-conducting bar of length b along the x-axis with one end at x = 0 held at temperature To and the other end at x = b held attemperature - To. The lateral surface of the bar is thermally insulated. Find the temperature distribution at all times if initially it is given by 2To

(a) T(O, x) = - b x

+ To,

2 To 2 (b) T(O, x) = -z;zx

To

+ To,

+ To,

(e) T(O, x) = -,;x

(d) T(O,x) = Tocosej;x),

0:" x:" b.

where where where

0:" x :" b.

0:" x < b.

where

O:"x:"b.

Hint: The solution corresponding to the zero eigenvalue is essential and cannot be excluded.

19.9. Determine T(x, y, r) for the rectangular plate of Example 19.2.3 if initially the lower left quarter is held at To and the rest of the plate is held at T = O. 19.10. All sides of the plate of Example 19.2.3 are held at T = O. Find the temperature distribution for all time if the initial temperature distribution is given by x< 4 ~a and < ~b To if "a 4<4Ib -<-y 4 ' otherwise.

(a) T(x, y, 0) = { 0

To

(b) T(x, y, 0) = -xy,

ab

To (e) T(x, y, 0) = -x,

a

where where

0:" x < a and 0 :" y < b. 0:" x < a and 0 < y < b.

19.11. Repeat Example 19.2.3 with the temperatures of the sides equal to Tl, T2, T3, and T4. Hint: You must include solutions corresponding to the zero eigenvalue.

19.12. A string of length a is fixed at the left end, and the right end moves with displacement A sinmt. Find 1{f(x, t) and a consistent set of initial conditions for the displacement and the velocity.

(

19.5 PROBLEMS

547

19.13. Find the equation for a vibrating rectangular membrane with sides of lengths a and b rigidly fastened on all sides. For a = b, show that a given mode

frequency may havemorethan one solution. 19.14. Repeat Example 19.3.1 if the can has semi-infinite length, the lateral surface is grounded, and: (a) the base is held at the potential V(p,
Vo a

Vo (c) V = -x. a

(b) V = -yo

Hint: Use the integral identity

f zV+IJv(z) dz =

Vo (d) V = '2xy. a

Zv+1 Jv+1 (z).

19.15. Find the steady-state temperature distribution T (p, tp, z) in a semi-infinite solid cylinder of radius a if the temperature distribution of the base is f(p,
(a) -yo a

To (b) -x.

a

To

(c) '2xy. a

(d) To.

19.18. Find the temperature of a circular conducting plate of radius a at all points of its surface for all time t > 0, assuming that its edge is held at T = 0 and initially its surface from the ceuter to a/2 is in contact with a heat bath of temperature To. 19.19. Fiud the potential of a cylindrical conducting can of radius a and height h whose top is held at a constant potential Vo while the rest is grounded. 19.20. Find the modes and the corresponding fields of a cylindrical resonant cavity of length L and radius a. Discuss the lowest TM mode. 19.21. Two identical long conducting half-cylindrical shells (cross sectious are half-circles) ofradius a are glued together in such a way that they are insulated from one another. One half-cylinder is held at potential Vo and the other is grouuded. Find the potential at any point inside the resulting cylinder. Hint: Separate Laplace's

equation in two dimensions.

(

548

19. STURM-LIOUVILLE SYSTEMS: EXAMPLES

19.22. A linear charge distribution of uniform dcnsity x extends along the z-axis from z = -b to z = b. Show that the electrostatic potential at any point r > b is given by
(b/r)Zk+l I: PZk(COSe). k~O 2k+ I 00

Hiut: Consider a point on the z-axis at a distance r > b from the origin. Solve the simple problem by integration and compare the result with the infinite series to obtain the unknown coefficients.

19.23. The upper half of a heat-conducting sphere of radius a has T = 100; the . lower halfis maintained at T = -100. The whole sphere is inside an infinitely large mass of heat -conducting material. Find the steady-state temperature distribution inside and outside the sphere. 19.24. Find the steady-state temperature distribution inside a sphere of radius a when the surface temperature is given by: (a) To cos z

e.

(d) To (cos e- cos

3

e).

(b) To cos" e.

(c) Tol cos e].

e.

(f) To sin4 e.

(e) To sm2

19.25. Find the electrostatic potential both inside and outside a conducting sphere of radius a when the sphere is maintained at a potential given by (a) Vo(cose - 3sinz e ) .

(c) {

Vo cos

o

e

for the upper hemisphere, .

for the lower hemisphere.

19.26. Find the steady-state temperature distribution inside a solid hemisphere of radius a if the curved surface is held at To and the flat surface at T = O. Hint: Imagine completing the sphere and maintaining the lower hemisphere at a temperature such that the overall surface temperature distribution is an odd function about e = n /2. 19.27. Find the steady-state temperature distribution in a spherical shell of inner radius R 1 and outer radius Rz when the inner surface has a temperature Tl and the outer surface a temperature Tz-

Additional Reading 1. Jackson, J. Classical Electrodynamics, 2nd ed., Wiley, 1975. The classic textbook on electromagnetism with many examples and problems on the solutions of Laplace's equation in different coordinate systems.

(

19.5 PROBLEMS

549

2. Mathews, J. and Walker, R. Mathematical Methods of Physics, 2nd ed., Benjamin, 1970. 3. Morse, P. and Feshbach, M. Methods ofTheoretical Physics, McGraw-Hill, 1953.

(

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

(

Part VI

_

Green's Functions

(

1 1

(

20

_

Green's Functions in One Dimension

Our treatment of differential equations, with the exception of SOLDEs with constant coefficients, did not consider inhomogeneous equations. At this point, however, we can put into use one of the most elegant pieces of machinery in higher mathematics, Green's functions, to solve inhomogeneous differential equations. This chapter addresses Green's functions in one dimension, that is, Green's functions of ordinary differential equations. Consider the ODE Lx[u] = f(x) where Lx is a linear differential operator. In the abstract Dirac notation this can be formally written as L lu) = If}. If L has an inverse L- 1 sa G, the solution can be formally written as lu) = L-I If} = G If}. Multiplying this by (xl and inserting 1 = dy Iy) w(y) (yl between G and If) gives

f

u(x) =

f

where the integration is over the range of definition of the functions involved. Once we know G(x, y), Equation (20.1) gives the solution u(x) in an integral form. But how do we find G(x, y)? Sandwiching both sides of LG = 1 between (xl and Iy) and using 1 = dx'lx'} w(x') (x'i between L and G yields dx'L(x, x')w(x')G(x', y) = (xl y) = 8(x - y)jw(x) if we use Equation (6.3). In particular, if L is a local differential operator (see Section 16.1), then L(x, x') = [8(x - x')jw(x)]Lx, and we obtain

f differential equation forGreen's function

Green's function

(20.1)

dyG(x, y)w(y)f(y),

f

8(x - y) LxG(x, y) = --'C....,-7"'w(x)

or

LxG(x, y) = 8(x - y),

(20.2)

where the second equation makes the frequently used assumption that w(x) = 1. G(x, y) is called the Green's function (GF) for the differential operator (DO) Lx.

(

554

20. GREEN'S FUNCTIONS IN ONE OIMENSION

As discussed iu Chapters 16 and 18, Lx might uot be defined for all functions on R Moreover, a complete specification of Lx reqnires some initial (or boundary) conditions. Therefore, we expect G(x, y) to depend on such initial conditions as well. We note that when Lx is applied to (20.1), we get Lxu(x) =

f

dy[Lx(G(x, y)]w(y)f(y) =

f

dy 8(x - y) w(y)f(y) = f(x), w(x)

indicating that u(x) is indeed a solution of the original ODE. Equation (20.2), involving the generalized function 8(x - y) (or disttibution in the language of Chapter 6), is meaningful only in the same context. Thus, we treat G(x, y) not as an ordinary function but as a distribution. Finally, (20.1) is assumed to hold for an arbitrary (well-behaved) function f.

20.1 Calculation of Some Green's Functions This section presents some examples of calculating G(x, y) for very simple DOs. Later we will see how to obtain Green's functions for a general second-order linear differential operator. Although the complete specification of GFs requires boundary conditions, we shall introduce unspecified constants in some of the examples below, and calculate some indefinite GFs.

20.1.1. Example. Let us find the GF for the simplestDO, Lx = d/dx. Weneed to find a distribution suchthat its derivative is theDirac deltafunction: 1 G'(x, y) = cS (x - y). In Chapter 6, we encountered such a distribution-the step function B(x - y). Thus, G(x, y) = O(x - y) + aryl, where aryl is the "constant" of integration. III The example above did not include a boundary (or initial) condition. Let us see how boundary conditions affect the resulting GF.

20.1.2. Example. Let ussolve u'(x) = f(x) where x E [0,00) andu(O) = O. A general solution of this DE is given by Equation (20.1) andthe preceding example: u(x)

= fooo O(x -

y)f(y)dy + foOO a(y)f(y)dy.

The factor8(x - y) in the firstterm on the RHS chops off the integralatx: u(x) = fox f(y)dy

+ fooo a(y)f(y)dy.

The Be gives0 = u(O) = 0+ Jooo a(y)f(y)dy. The only waythat this can be satisfied for arbitrary fey) is for aryl to be zero. Thus, G(x, y) = O(x - y), and u(x) =

foOO O(x -

y)f(y)dy = fox f(y)dy.

1Hereandelsewhere in thischapter, a primeovera GF indicates differentiation withrespectto its first argument.

(

20.1 CALCULATION OF SOME GREEN'S FUNCTIONS

555

Thisis killingafly witha sledgehammer! Wecouldhaveobtained theresult by a simple integration. However, theroundabout way outlined hereillustrates someimportant features of GFsthat will be discussed later. The Be introduced hereis very special. What happens if it is changed to u(O) = a? Problem 20.1 answers lbat.

III

20.1.3. Example. A more complicated DO is Lx = d 2/ dx 2. Let us find its indefinite GF. To do so, we integrate GI/(x, y) = 8(x - y) ouce wilbrespect to x to obtain -!xG(x, y) = O(x - y) + a(y). A secoud integration yields G(x, y) = f dxO(x - y) + xa(y) + ~(y), where a and 11 arearbitrary functions andtheintegral is anindefinite integral to beevaluated next.

Let Q(x, y) be the primitive of O(x - y); lbat is,

-dQ"( =u x-y) =

{I

dx

0

if x > y, if x < y.

(20.3)

The solution to this equation is Q(x, y) = {x +a(y) bey)

if x > y, if x < y.

Note that we have not defined Q(x, y) atx = y. It will become c1earbelow lbat Q(x, y) is continuous at x = y. It is convenient to write Q (x, y) as (20.4)

Q(x, y) = [x +a(y)]e(x - y) +b(y)e(y -x).

To specify aryl and bey) further, we differentiate (20.4) and compare it wilb (20.3):

dQ

+ [x + a(y)]8(x - y) - b(y)8(x y) + [x - bey) + a(y)]8(x - y),

dx = O(x - y)

=O(x -

y) (20.5)

where we have used

d

dx O(x - y)

d

= - dx O(y -

x)

= 8(x -

y).

For Equation (20.5) to agree wilb (20.3), we must bave [x - bey) + a(y)]8(x - y) = 0, which,uponintegration overx, yields a(y) - b(y) = -y. Substituting thisin theexpression for Q(x, y) gives Q(x, y) = (x - y)e(x - y)

But O(x)

+ Or-x)

+ b(y)[O(x -

+ O(y -

x)].

+ bey). It follows, among = y. We cannow write

= I; therefore, Q(x, y) = (x - y)e(x - y)

other things, that n (x, y) is continuous at x G(x, y) = (x - y)e(x - y)

where fJ(y) =

y)

~(y)

+ bey).

+ xa(y) + fJ(y), III

The GF in the example above has two arbitrary functions, Ol(y) and fJ (y), which are the result of underspecification of Lx: A full specification of Lx requires Bes, as the following example shows.

(

556

20. GREEN'S FUNCTIONS IN ONE OIMENSION

20.1.4.Example. Let us calculate the GF of Lx[u]

= u"(x) = f(x) subject to the BC u(a) = u(b) = 0 where [a, b] is the iuterval on which Lx is defined.Example 20.1.3 gives us the (indefinite) GF for Lx.Using that,we can write

u(x) = =

l

b

(x - y)e(x - y)f(y) dy + x

LX (x -

+x

l

a(y)f(y) dy

+

y)f(y) dy

l

b

a(y)f(y) dy +

b

a(y)f(y) dy

+

l

l

b fJ(y)f(y) dy

b

fJ(y)f(y) dy.

Applyiug the BCs yields

l

0= u(a) = a 0= u(b) =

l

b

b

(b - y)f(y) dy

l

+b

b

fJ(y)f(y) dy,

l

b

a(y)f(y) dy

+

t

fJ(y)f(y) dy.

(20.6)

From these two relations it is possibleto determine ex (y) and fJ (y): Substitute for the last iutegral on the RHS of the second equation of (20.6) from the first equation and get o = J/;[b - y + ba(y) - aa(y)]f(y) dy. Siuce this must hold for arbitrary fey), we conclude that b- y

+ (b -

a)a(y)

= 0 =>

a(y)

b-y

= ---. b-a

Substituting for a (y) iu the first equation of (20.6) and noting that the result holds for arbitrary f, we obtain fJ(y) = a(b-y)/(b-a). Insertionofa(y) andfJ(y) iutheexpression for G(x, y) obtaiued iu Example 20.1.3 gives G(x, y)

= (x -

y)e(x - y)

It is striking that G(a, y)

+ (x -

= (a -

G(b, y) = (b - y)e(b - y)

+ (b -

y-b a) b _ a

y)e(a - y)

where

a:::; x and y ::s: b.

= 0 (because a -

y

~

0), and

y-b a) b _ a = 0

because e(b - y) = 1 for all y ~ b [recall that x and y lie iu the iuterval (a, b)J. These two equations reveal the important fact thatas a function of x, G(x, y) satisfies the same (homogeneous) Be asthesolutionof the DE.Thisis a generalproperty that will bediscussed later. III

In all the preceding examples, the BCs were very simple, Specifically, the value of the solution and/or its derivative at the boundary points was zero. What if the BCs are not so simple? In particular, how can we handle a case where u(a) [or u'(a)] and u(b) [or u'(b)] are nonzero? Consider a general (second-order) differential operator Lx and the differential equation Lx[u] = J(x) subjecttotheBCsu(a) = at andu(b) = bl. We claim that we can reduce this system to the case where u(a) = u(b) = O.RecallfromChapter 13 that the most general solution to such a DE is of the form u = Uk + u, where Uk,

(

20.2 FORMAL CONSIDERATIONS

557

the solution to the homogeneous equation, satisfies Lx[Uh] = 0 and contains the arbitrary parameters inherent in solutions of differential equations. For instance, if the linearly independent solutions are v and w, then Uh(X) = C, v(x) + C2W(X) and Ui is any solution of the inhomogeneous DE. If we demand that uh(a) = a, and uh(b) = b, then u, satisfies the system Lx[uil = f(x),

Uj(a)

= ui(b) = 0,

which is ofthe type discussed in the preceding examples. Since Lx is a SOLDO, we can put all the machinery of Chapter 13 to work to obtain v (x), w (x), and therefore Uh(X). The problem then reduces to a DE for which the BCs are homogeneous; that is, the value of the solution and/or its derivative is zero at the boundary points.

20.1.5.Example. Let us asswne that Lx = d 2/dx 2 Calculation of Uh is trivial: d2uh Lx[Uh] = 0 ~ -2- = 0 ~ Uh(X) = c'x dx

+ C2·

To evaluate C, and C2, we impose the BCs Uh(a) = a, and uh(b) = b,: C,a C,b

+ C2 = + C2 =

at. b,.

This gives C, = (b, - a,)/(b - a) and C2 = (a,b - ab,)/(b - a). The inhomogeneous equation defines a problem identical to that of Example 20.1.4. Thus, we can immediately write Ui(X) = G(x, y)f(y) dy, where G(x, y) is as given in thatexample.Thus,the generalsolutionis

I:

u(x)

hI - at = ~x + albb -_ aabl +

l

x a (x - y)f(y) dy

alb

+ xb -_ a

a (y - b)f(y) dylll

Example 20.1.5 shows that an inhomogeneous DE with inhomogeneous BCs can be separated into two DEs, one homogeneous with inhomogeneous BCs and the other inhomogeneous with homogeneous BCs, the latter being appropriate for the OF. Futthermore, all the foregoing examples indicate that solutions of DEs can be succinctly written in terms of OFs that automatically incorporate the BCs as long as the BCs are homogeneous. Can a OF also give the solution to a DE with inhomogeneons BCs?

20.2 Formal Considerations The discussion and examples of the preceding section hint at the power of Green's functions. The elegance of such a function becomes apparent from the realization that it contains all the information about the solutions of a DE for any type of BCs, as we are about to show. Since OFs are inverses of DOs, let us briefly reexamine the inverse of an operator, which is closely tied to its specttum. The question as

(

558

20. GREEN'S FUNCTIONS IN ONE OIMENSION

to whether or not an operator A in a finite-dimensional vector space is invertible is succinctly answered by the value of its determinant: A is invertible if and only if det A t= O. In fact, as we saw at the beginning of Chapter 16, one translates the abstract operator equation A lu) = [u) into a matrix equation Au = v and reduces the question to that of the inverse of a matrix. This matrix takes on an especially simpleform when A is diagonal,that is, when Aij = AiBij. For this special situation we have for i = 1,2, ... , N

(no sum over i).

(20.7)

This equation has a nnique solution (for arbitrary Vi) if and only if Ai t= 0 for all i, In that case Ui = Vi/Ai for i = 1,2, ... , N. In particnlar, if Vi = 0 for all i, that is, when Equation (20.7) is homogeneous, the unique solution is the trivial solution. On the other hand, when some of the Ai are zero, there may be no solution to (20.7), but the homogeneous equation has a nontrivial solution (Ui need not be zero). Recalling (from Chapter 3) that an operator is invertible if and only if none of its eigenvaluesis zero, we have the following: 20.2.1. Proposition. The operator A E J:., (V) is invertible if and only if thehomo-

geneous equation A Iu) = 0 has no nontrivial solutions. In infinite-dimensional (Hilbert) spaces there is no determinant. How can we tell whether or not an operator in a Hilbert space is invertible? The exploitation of the connection between invertibility and eigenvalues has led to Proposition 20.2.1, which can be generalized to an operator acting on any vector space, finite or infinite. Consider the equation A lu) = 0 in a Hilbert space 1(. In general, neither the domain nor the range of A is the whole of 1(. If A is invertible, then the only solution to the equation A lu) = 0 is lu} = O. Conversely, assuming that the equation has no nontrivial solution implies that the null space of A consists of only the zero vector.Thus,

This showsthat A is injective (one-to-one),i.e., A is a bijectivelinear mappingfrom the domain of A, 1J(A), onto the range of A. Therefore, A must have an inverse. The foregoing discussion can be expressed as follows. If A lu) = 0, then (by the definition of eigenvectors) A = 0 is an eigenvalue of A if and only if lu) t= o. Thus, if A Iu) = 0 has no nontrivial solution, then zero cannot be an eigenvalueof A. This can also be stated as follows: 20.2.2. Theorem. An operator A on a Hilbertspace has an inverse if and only if A = 0 is not an eigenvalue of A. Green's functions are inverses of differential operators. Therefore, it is important to have a clear understanding of the DOs. An nth-order linear differential operator (NOLDO)satisfiesthe followingtheorem (for a proof, see [Birk 78, Chapter 6]).

(

559

20.2 FORMAL CONSIOERATIONS

20.2.3. Theorem. Let d" d n- I P,,(x)-dn + Pn-I(X)-d1+'" x x":

Lx =

d

+ PI (x)-dX + po(x)

(20.8)

where p,,(x) ¢ 0 in [a, b]. Let Xo E [a, b] and let {YklZ~1 be given numbers and f(x) a given piecewise continuous function on [a, b]. Then the initial value problem (IVP) Lx[u] = f

for x E [a, b],

u(xo) = YI, u'(xo) = J!2, ... , uC,,-I\xO) = Yn

(20.9)

has one and only one solution. initial value problem

This is simply the existence and uniqueness theorem for a NOLDE. Equation (20.9) is referred to as the IVPwith data (f(x); YI, ... , YnJ. This Iheoremis used to define Lx. Part of Ihat definition are the BCs Ihat the solutions to Lx must satisfy. A particularly important BC is the homogeneous one in which YI = J!2 = ... = Yn = O. In such a case it can he shown (see Problem 20.3) Ihat the only nontrivial solution of the homogeneous DE Lx[u] = 0 is u ss O. Theorem 20.2.2 then tells us that Lx is invertible; that is, there is a unique operator G such that LG = 1. The "components" version of Ihis last relation is part of the content of the

nexttheorem. 20.2.4. Theorem. The DO Lx of Equation (20.8) associated with the IVP with data (f(x); 0, 0, ... , OJ is invertible; that is, there exists afunction G(x, y) such that .

LxG(x, y) =

8(x-y)

w(x)

.

The importance of homogeneous BCs can now be appreciated. Theorem 20.2.4 is the reason why we had to impose homogeneous BCs to obtain the GF in all the examples of the previous section. The BCs in (20.9) clearly are not the only ones Ihat can be used. The most general linear BCs encountered in differential operator Iheory are

RI[U] sa al1u(a) + R2[U] == a2Iu(a) +

+ alnu(n-I)(a) + fJl1u(b) + + a2nu(n-I)(a) + fJ2IU(b) +

+ fJlnuCn-I)(b) = YI, + fJ2nU(n-l) (b) = J!2, (20.10)

Rn[u] == anlu(a)

+ ... + annuC,,-I)(a) + fJ"lu(b) + ... + fJnnuCn-l\b)

= Yn·

The n row vectors {(ail, ... , ain, fJil, ... , fJin)}7~1 are assumed to be independent (in particular, no row is identical to zero). We refer to Ri as boundary functionals because for each (sufficiently smoolh) function u, Ihey give a number Yi. The

(

560

20. GREEN'S FUNCTIONS IN ONE DIMENSION

boundary functionals and boundary value problem

DO of (20.8) and the BCs of (20.10) together form a boundary valne problem (BVP). The DE Lx[u] = f subject to the BCs of (20.10) is a BVP with data (f(x); YI, ... , Ynl. We note that the Ri are linear; that is, and

completely homogeneous problem

adjoint ofa differential operator

Since Lx is also linear, we conclude that the superposition principle applies to the system consisting of Lx[u] = f and the BCs of (20.10), which is sometimes denoted by (L; RI, ... , Rn). If u satisfies the BVP with data {f; YI, ... , Ynl and v satisfies the BVP with data Is: 1."1, ... , /Lnl, then au + pv satisfies the BVP with data laf + pg; an + P/LI, , aYn + P/Lnl. It follows that if u and v both satisfy the BVPwith data {f; n, , Ylll, then u - v satisfies the BVP with data 10; 0, 0, ... , O], which is called the completely homogeneons problem. Unlike the IVP, the BVP with data 10; 0, 0, ... , O} may have a nontrivial solution. If the completely homogeneous problem has no noutrivial solution, then the BVP with data {f; YI, ... , Ynl has at most one solution (a solution exists for any set of data). On the other hand, if the completely homogeneous problem has nontrivial solutions, then the BVP with data {f; YI, ... , Yn1either has no solutions or has more than one solution (see [Stak 79, pp. 203-204]). Recall that when a differential (unbounded) operator Lx acts in a Hilbert space, such as.c~ (a, b), it acts only on its domain. In the context ofthe present discussion, this means that not all functions in.c~ (a, b) satisfy the BCs necessary for defining Lx. Thus, the functions for which the operator is defined (those that satisfy the BCs) form a subset of .c~(a, b), which we called the domain of Lx and denoted by 'D(Lx). From a formal standpoint it is important to distinguish among maps that have different domains. For instance, the Hilbert-Schmidt integral operators, which are defined on a finite interval, are compact, while those defined on the entire real line are not. 20.2.5. Definition. Let Lx be the DO of Equation (20.8). Suppose there exists a DO Lt, with the property that

d w Iv* (Lx [u]) - u(Lt[v])*l = -Q[u, v*] dx conjunct

for u, v E 'D(Lx) n'D(Lt),

where Q[u, v*l, called the conjunct ofthe functions u and v, depends on u, v, and their derivatives oforder up to n - I. The DO Lt is then called the formal adjoint of Lx· IfLt = Lx (without regard to the Bes imposed on their solutions), then Lx is said to beformally self-adjoint. If'D(Lt) ::J 'D(L x) and Lt = Lx on 'D(Lx), then Lx is said to be hermitian. If 'D(Lb 'D (Lx) and Lt Lx, then Lx is said to be self-a4joint.

=

generalized Green's identity

Ri[au] = aRi[U].

=

The relation given in the definition above involving the conjunct is a generalization of the Lagrange identity and can also be written in integral form:

(

20.2 FORMAL CONSIOERATIONS

l

b

-l dxw{u(L~[V])*} b

dxw{v*(Lx[u])}

= Q[u,

v*]I~

561

(20.11)

This form is sometimes called the generalized Green's identity.

George Green (17937-1841) was not appreciated in his lifetime. His date of birth is unknown (however, it is known that he was baptized on 14 July 1793), and no portrait of him survives. He left school,after only one year's attendance,to work in his father's bakery. When the father opened a windmill in Nottingham, the boy used an upper room as a study in which he taught himself physics and mathematics from library books. In 1828, when he was thirty-five years old, he published his most important work, An Essay on the Application of Mathematical Analysis to the Theory of Electricity and Magnetism at his own expense. In it Green apologized for any shortcomings in the paper due to his minimal formal education or the limited resources available to him, the latter being apparent in the few previous works he cited. The introduction explained the importance Green placed on the "potential" function. The body of the paper generalizes this idea to electricity and magnetism. In addition to the physics ofelectricity and magnetism, Green's first paper also contained the monumental mathematical contributions for which he is now famous: The relationship between surface and volume integrals we now call Green's theorem, and the Green'sfunction, a ubiquitous solution to partial differential equations in almost every area of physics. With little appreciation for the future impact of this work, one of Green's contemporaries declared the publication "a complete failure." The "Essay", which received little notice because of poor circulation, was saved by Lord Kelvin, who tracked it down in a German

journal. When his father died in 1829, some of George's friends urged him to seek a college education. After four years of self-study, during which he closed the gaps in his elementary education, Green was admitted to Caius College of Cambridge University at the age of 40, from which he graduated four years later after a disappointing performance on his final examinations. Later, however, he was appointed Perce Fellow of Caius College. Two years after his appointment he died, and his famous 1828 paper was republished, this timereaching a much wider audience. This paper has been described as "the beginning of mathematical physics in England." He published only ten mathematical works. In 1833 he wrote three further papers. Two on electricity were puhlished by the Cambridge Philosophical Society. One on hydrodynamics was published by the Royal Society of Edinburgh (of which he was a Fellow) in 1836. He also had two papers on hydrodynamics (in particular wave motion in canals), two papers on reflection and refraction of light, and two papers on reflection and refraction of sound published in Cambridge. In 1923 the Green windmill was partially restored by a local businessman as a gesture of tribute to Green. Einstein came to pay homage. Then a fire in 1947 destroyed the renovations. Thirty years later the idea of a memorial was once again mooted, and sufficient money was raised to purchase the mill and present it to the sympathetic Nottingham City Council. In 1980 the George Green Memorial Appeal was launcbed to secure $20,000 to get the sails

(

562

20. GREEN'S FUNCTIONS IN ONE OIMENSION

turning again and the machinery working once more. Today, Green's restored mill stands as a mathematics museum in Nottingham.

20.2.1 Second-Order Linear DOs Since second-order linear differential operators (SOLDOs) are sufficiently general for most physical applications, we will concentrate on them. Because homogeneous BCs are important in constructing Green's functions, let us first consider BCs of the form

+ "12U' (a) + fJl1u(b) + fJ12u' (b) sa "21u(a) + "22U' (a) + f!21u(b) + fJ22u' (b)

Rl [u] '" "l1u(a)

= 0,

R2[ul

= 0,

(20.12) where it is assumed, as usual, that ("11, "12, fJl1, fJ12) and ("21, "22, fJ21, fJ22) are linearly independent. If we define the inner product as an integral with weight w, Equation (20.11) can be formally written as (vi L lu) = (ul Lt Iv)*

+ Q[u, v*]I~.

This would coincide with the usual definition of the adjoint if the surface term vanishes, that is, if Q[u, v*llx=b = Q[u, v*]lx=a.

(20.13)

For this to happen, we need to impose BCs on v. To find these BCs, let us rewrite Equation (20.12) in a more compact form. Linear independence of the two row vectors of coefficients implies that the 2 x 4 matrix of coefficients has rank two. This means that the 2 x 4 matrix has an invertible 2 x 2 submatrix. By rearranging the terms in Equation (20.12) if necessary, we can assume that the second of the two 2 x 2 subrnatrices is invertible. The homogeneous BCs can then be conveniently written as

1 R[u] = (Rl[U R2[U]) = (A

B)

(UUba)

= AUa + BUb = 0 ,

(20.14)

where B '" (fJl1 fJ21

and B is invertible.

fJ12), fJ22

u (a ) )

ua sa ( u'(a) ,

Ub'"

u (b) ) ( u'(b) ,

(

20.2 FORMAL CONSIOERATIONS

563

The most general form of the conjnnct for a SOLDO is

+ q12(X)U(X)v'*(x) + q21 (x)u' (x)v*(x) + q22(X)U'(x)v'*(x),

Q[u, v*](x) '= ql1 (x)u(x)v*(x)

which can he written in matrix form as Q[u, v*](x) = u~axv;

ax = (ql1 (x)

where

Q21(X)

qJ2(X)), Q22(X)

(20.15)

and Ux and v; have similar definitions as Ua and Ub ahove. The vanishing of the surface termbecomes (20.16)

We need to translate this equation into a condition on v* alone/' This is accomplished hy solving for two of the four quantities u(a), u'(a), u(b), and u'(b) in terms of the other two, suhstituting the result in Equation (20.16), and setting the coefficients of the other two equal to zero. Let us assume, as before, that the suhmatrix B is invertible, i.e., u(b) and u'(b) are expressible in terms of u(a) and u'(a). Then Ub = -B-1Aua, or ub = _u~At(Bt)-I, and we obtain -u~At(Bt)-labVb = u~aav~

=}

u~ [At(Bt)-labVb +aav~] = 0,

andthe conditionon v* becomes (20.17) We see that all factors of u have disappeared, as they should. The expanded version of the BCs on v* are written as

+ u12v'*(a) + ql1v*(b) + q12v'*(b) = B2[V*] sa u2Iv*(a) + U22 v'* (a) + Q2IV*(b) + Q22V'*(b) = BI[v*] es ul1v*(a)

0, O. (20.18)

adjoint boundary conditions

These homogeneous BCs are said to be adjoint to those of (20.12). Because of the difference between BCs and their adjoints, the domain of a differential operator need not be the sarne as that of its adjoint. 20.2.6. Example. Let Lx = d 2/ dx 2 with the homogeneous BC, RI[U] = au(a) - u' (a)

=0

and

R2[U]

= f3u(b) -

u'(b)

= O.

(20.19)

Wewant to calculate Q[u, v*] and the adjointBC, for v. By repeatedintegration by parts [or by using Equation (13.23)], we obtain Q[u, u"] vanish, we musthave

= u'v* - uv'*. For the surface termto

u'(a)v*(a) - u(a)v'*(a) = u'(b)v*(b) - u(b)v'*(b). 2Theboundary conditions on v* should not depend on thechoice of u.

564

20. GREEN'S FUNCTIONS IN ONE DIMENSION

Substitutingfrom (20.19)in this equation, we get u(a)[av*(a) - v'*(a)]

= u(b)[{lv*(b) -

v'*(b)],

which holds for arbitrary u if and only if Bl[V*]

= av*(a) -

v'*(a)

=0

and

B2[V*]

= {lv*(b) -

v'*(b)

= O. (20.20)

Thisis a special case, in whichtheadjoint Bes arethesameastheoriginal Bes (substitute u for v* to see this). To see that theoriginal Bes andtheir adjoints neednot be the same, we consider Rdu]

= u'(a) -

au (b)

=0

and

R2[U]

= {lu(a) -

u'(b)

= 0,

(20.21)

from which we obtain u(a)[{lv*(b) + v'*(a)] = u(b)[av*(a) Bl[V*]

= av*(a) + v'*(b) = 0

and

+ v'*(b)]. Thus, BZ[v*] = {lv*(b) + v'*(a) = 0, (20.22)

mixed and unmixed BCs

which is not the sameas (20.21).Boundaryconditionssuch as those in (20.19) and (20.20), in whicheachequation contains thefunction andits derivative evaluated: atthesamepoint, are called unmixed BCs. On the other hand, (20.21) and (20.22) are mixed BCs. III

20.2.2 Self-Adjoint SOLDOs ill Chapter 13, we showed that a SOLDO satisfies the generalized Green's identity with w(x) = I. ill fact, since u and v are real, Eqnation (13.24) is identical to (20.11) if we set w = 1 and Q[u, v] = pzvu' - (P2V)'U

+ PIUV.

(20.23)

Also, we have seen that any SOLDO can be made (formally) self-adjoint. Thus, let us consider the formally self-adjoint SOLDO Lx = L% =

:x (p :x) +q

where both p(x) and q(x) are real functions and the inner product is defined with weight w = I. lf we are interested in formally self-adjoint operators with respect to a general weight w > 0, we can construct them as follows. We first note that if Lx is formally self-adjoint with respect to a weight of unity, then (l/w)L x is selfadjoint with respect to weight w. Next, we note that Lx is formally self-adjoint for all functions q, in particular, for wq. Now we define L(w)

x

=

~ (p~) +qw

dx

dx

w

and note that Li therefore

)

is formally self-adjoint with respect to a weight of unity, and

lw l-d d) Lx= -Li)= - (P-

w

wdx

dx

+q

(20.24)

20.3 GREEN'S FUNCTIONS FOR SOlDOS

565

is formally self-adjoint with respect to weight w(x) > O. For SOLDOs that are formally self-adjoint with respect to weight w, the conjnnct given in (20.23) reduces to

Q[u, v] = p(x)w(x)(vu' - uv').

(20.25)

Thus, the surface term in the generalized Green's identity vanishes if and only if

p(b)w(b)[v(b)u'(b) - u(b)v'(b)] = p(a)w(a)[v(a)u'(a) - u(a)v'(a)]. (20.26) common types of boundary conditions for aSOlDE

The DO becomes self-adjoint if u and v satisfy Equation (20.26) as well as the same BCs. It can easily be shown that the following four types of BCs on u(x) assure the validity of Equation (20.26) and therefore define a self-adjoint operator Lx given by (20.24):

= u(b) = 0 The Neumann BCs: u'(a) = u'(b) = 0

I. The Dirichlet BCs: u(a) 2.

= fJu(b) Periodic BCs: u(a) = u(b) and u'(a) = u'(b)

3. General unmixed BCs: au(a) - u'(a) 4.

u'(b)

=0

20.3 Green's Functions for SOLDOs We are now in a position to find the Green's function for a SOLDO. First, note that a complete specification of Lx requires not only knowledge of po(x), Pt (x), and p2(x)-its coefficient functions-but also knowledge of the BCs iroposed on the solutions. The most general BCs for a SOLDO are of the type given in Equation (20.10) with n = 2. Thus, to specify Lx uniqnely, we consider the system (L; RI, R2) with data (f; VI, )12). This system defioes a unique BVP:

d 2u du Lx[u] = P2(X)dx 2 + PI (x) dx

+ po(x)u = f(x),

Ri[U]

= Vi,

i

= 1,2. (20.27)

A necessary condition for Lx to be invertible is that the homogeneous DE Lx[u] = 0 have only the trivial solution u = O. For u = 0 to be the only solntion, it must be a solntion. This means that it must meet all the conditions in Equation (20.27). In particular, since Ri are linear functionals of u, we must have Ri [0] = O. This can be stated as follows: 20.3.1. Lemma. A necessary condition for a second-order linear DO to be invertible is for its associated BCs to be homogeneous? 3The lemma applies to all linear DOs, not just second order ones.

566

20. GREEN'S FUNCTIONS IN ONE OIMENSION

Thus, to study Green's functions we must restrict ourselves to problems with homogeneous BCs. This at first may seem restrictive, since not all problems have homogeneons BCs. Can we solve the others by the Green's function method? The answer is yes, as will be shown later in this chapter. The above discussion clearly indicates that the Green's function of Lx> being its "inverse," is defined only if we consider the system (L; Rj, Rz) with data (I; 0, 0). If the Green's function exists, it must satisfy the DE of Theorem 20.2.4, in which Lx acts on G(x, y). But part of the definition of Lx are the BCs imposed on the solutions. Thus, if the LHS of the DE is to make any sense, G(x, y) must also satisfy those sarae BCs. We therefore make the following definition: formal definition of Green's function

20.3.2. Definition. The Green's function of a DO Lx is a function G(x, y) that satisfies both the DE 8(x - y)

LxG(x, y) =

w(x)

and, as afunction of x, the homogeneous Bes Ri[G] = Ri are defined as in Equation (20.12).

Ofor i

= 1,2 where the

It is convenient to study the Green's fumction for the adjoint of Lx simultaneously. Denoting this by g(x, y), we have t

8(x - y)

Lxg(x, y) =

adjoint Green's function

w(x)

,

Bi[g] = 0,

for i = 1,2,

(20.28)

where Bi are the boundary fumctionals adjoint to Ri and given in Equation (20.18). The function g(x, y) is known as the adjoint Green's function associated with the DE of (20.27). We can now use (20.27) and (20.28) to find the solutions to

Lx[u] = f(x),

Ri[U] = 0

for i = 1,2,

L1[v] = hex),

Bi[V*] = 0

for i = 1,2.

(20.29)

With v(x) = g(x, y) in Equation (20.11 )-whose RHS is assumed to he zero-we getJ: wg*(x, y)Lx[u] dx = J: wu(x)(d[gj)*dx. Using (20.28) on the RHSand (20.29) on the LHS, we obtain

u(y) =

l

b

g*(x, y)w(x)f(x) dx.

Similarly, with u(x) = G(x, y), Equation (20.11) gives

v*(y) =

l

b

G(x, y)w(x)h*(x) dx ;

20.3 GREEN'S FUNCTIONS FOR SOLO OS

567

or, since w (x) is a (positive) real function,

1 b

v(y) =

G*(x, y)w(x)h(x)dx.

These equations for u(y) and v(y) are not what we expect [see, for instance, Equatiou (20.1)]. However, if we take into account certain properties of Green's functions that we will discuss next, these equations become plausible.

20.3.1 Properties of Green's functions Let us rewrite the generalized Green's identity [Equation (20.11)], with the RHS equalto zero, as

1 b

Green's identity

1 b

dtw(t){v*(t)(Lt[u])) =

dtw(t){u(t)(L;[vj)*j.

(20.30)

This is sometimes called Green's identity. Substituting G(t, y) for u(t) and g(t, x) for vet) gives b

1

dtw(t)g*(t, x)

a

8(t

y) = wet)

1 b

. 8(t

dtw(t)G(t, y)

a

x) , wet)

or g* (y, x) = G(x, y). A consequence of this identity is that

20.3.3. Box. G(x, y) must satisfy the adjoint Bes with respect to its second argument.

If for the time being we assume that the Green's function associated with a system (L; RI, Rz) is unique, then, since for a self-adjoint differential operator, Lx and Ll are identical and u and v both satisfy the same BCs, we must have G(x, y) = g(x, y) or, using g*(y, x) = G(x, y), we get G(x, y) = G*(y, x). In particular, if the coefficient functions of Lx are all real, G (x, y) will be real, and we have G (x, y) = G (y, x). This means that G is a symmetric function ofits two arguments. The last property is related to the continuity of G(x, y) and its derivative at x = y. For a SOLDO, we have

aG + PI(x)-a + po(x)G = ax x

aZG LxG(x, y)= pz(x)-z

8(x - y) ()' W x

where PO, PI, and PZ are assumed to be real and continuous in the interval [a, b], and w(x) and pz(x) are assumed to be positive for all x E [a, b]. We multiply both sides of the DE by Jk(x) hex) = - - ) ' pz(x

where

Jk(x) '" exp

[l

a

x

PI (t) ] -(-) dt , PZ t

568

20. GREEN'S FUNCTIONS IN ONE OIMENSION

notingthatdfL/dx = (PI/P2)fL. This transforms the DE into

a[

a

ax fL(X) ax G(x, y)

]

+

PO(X)fL(x) fL(Y) P2(X) G(x, y) = P2(Y)W(Y) 8(x - y).

Integrating this equation gives

a

fL(X)ijG(x,y) x

+

1

x Po(t)fL(t)

a

()

P2 t

G(t,y)dt=

, fL(Y) () ()O(x-y)+a(y) P2 y w Y (20.31)

because the primitive of 8(x - y) is O(x - y). Here a(y) is the "constant" of integration. First consider the case where Po = 0, for which the Green's function will be denoted by Go(x, y). Then Eqnation (20.31) becomes fL(X)aa Go(x,y) = (;y)( )O(x - y) +aj(y), x P2YWY

which (since u, P2, and w are continuous on [a, b], andO(x-y) has a discontinuity only at x = y) indicates that aGo/ax is continuous everywhere on [a, b] except at x = y. Now divide the last equation by fL and integrate the result to get

Every term on the RHS is continuous except possibly the integral involving the O-function. However, that integral can be written as

/,X

x O(t - y) dt -'-...,....:--'-dt = O(x - y) -. a fL(t) y fL(t)

1

(20.32)

The O-function in front of the integral is needed to ensure that a ~ y ~ x as demanded by the LHS of Equation (20.32). The RHS of Equation (20.32) is continuous at x = y with limit being zero as x -+ y. Next, we write G(x, y) = Go(x, y) + H(x, y), and apply Lx to both sides. This gives 8(x - y) w(x) = =

(d

2

P2dx2

8(x - y) w(x)

d) Go +

+ PI dx

poGo

+ LxH(x, y)

+ poGo + LxH(x, y),

or P2H" + P1H' + PoH = -PoGo· The continuity of Go, po, PI, and P2 on [a, b] implies the continuity of H, because a discontinuity in H wonld entail a delta function discontinuity in d H / dx; which is impossible because there are no delta functions in the equation for H. Since both Go and H are continuous, G must also be continuous on [a, b].

20.3 GREEN'S FUNCTIONS FOR SOLOOS

569

We can now calcnlate the jnmp in aG/ax at x = y. We denote the jnmp as AG' (y) and define it as follows:

AG , (y)

. [aG == Inn -(x, y) I 10--+0

ax

x=y+€

aG - -(x, y)

ax

I

].

X=y-E

Dividing (20.31) by JL(x) and taking the above limit for all terms, we obtain

AG'(y)

+ lim

[

e....0

1 1 Y

I +' pO(t )JL (t ) G (t , y ) dt JL(Y+f) a pz(t)

1 JL(Y - f) JL(Y)

.

pz(y)w(y)!~

=

Y - '

a

Po(t)JL(t) G(t, y) dt] pz(t)

=1

~o

,-"-,

,-"-,

[O(+f) O(-f)] JL(Y + f) - JL(Y - f) .

The second term on the LHS is zero becanse all functions are continnons at y. The limit on the RHS is simply 1/JL(Y). We therefore obtain

AG' (y) =

1 pz(y)w(y)

(20.33)

20.3.2 Construction and Uniqueness of Green's Fnnctions We are now ina position to calcnlate the Green's function for a general SOLDO and show that it is unique. existence and uniqueness at GF tor a second order linear differential operator

20.3.4. Theorem. Consider the system (L; RI, Rz) with data (f; 0, 0), in which Lx is a somo. If the homogeneous DE Lx[u] = 0 has no nontrivial solution, then the GF associated with the given system exists and is unique. The solution of the system is

1 b

u(x) =

dyw(y)G(x, y)f(y)

and is also unique.

Proof The GF satisfies the DE Lx G(x, y) = 0 for all x E [a, b] except x = y. We thus divide [a, b] into two intervals, II = [a, y) and Iz = (y, b], and note that a general solution to the above homogeneous DE can be written as a linear combination of a basis of solutions, Uj and uz. Thus, we can write the solution of the DE as

Gii», y) = CjUj(x) +czuz(x)

for x E II

Gr(x,y) =dlul(x)+dzuz(x)

for x E

h

570

20. GREEN'S FUNCTIONS IN ONE OIMENSION

and define the GF as G(x, y) = {Gl(X, y)

o,i», y)

ifxEII, ifxEIz,

(20.34)

where CI, cz, di ; and dz are, in general, functions of y. To determine G(x, y) we must determine four unknowns. We also have fuur relations: the continuity of G, the jump in aGjax atx = y, and the two BCs RI[G] = Rz[G] = O. The continuity ofG gives

The jump of aGjax atx = y yields I Pz(y)w(y) .

= CJ -

Introducing bl to blu, f

b,uI

dl and bz

+ bzuz =

0,

+ bzuz =

I ---. pzw

f

= ci -

dz changes the two preceding equations

These equations have a unique solution iff det

C UZ)z I I

f

f

U

~

O.

But the determinant is simply the Wronskian of the two independent solutions and therefore cannot be zero. Thus, bl (y) and bz(y) are determined in terms ofu" u\' U2. pz. and w. We now define

u;.

hex, y) sa {bOI(y)U I (x)+ bz(y)uz(x)

if x E II, ifxE/z.

so that G(x, y) = hex, y)+dl(y)u, (x)+dz(y)uz(x). We have reduced the number of unknowns to two, d, and dz. Imposing the BCs gives two more relations: R,[Gj

= RI[h] +dIR,[uiJ +dzRI[uz] = 0,

Rz[G]

= Rz[h] + dIRz[uI] + dzRz[uz] = O.

Can we solve these equations and determine dl and dz uniquely? We can, if

20.3 GREEN'S FUNCTIONS FOR SOlDOS

571

It can be shown that this determinant is nonzero (see Problem 20.5). Having found the unique {bi, dil~=l' we can calculate ci uniquely, substitute all of them in Equation (20.34), and obtain the unique G(x, y). That u(x) is also unique can be shown similarly.

D

20.3.5. Example. Let us calculate the GFfor Lx = d 2/dx2 with BCs u(a) = u(b) = O. We note that Lx[u] = 0 with the given BCs has no nonnivial solution (verify this). Thns, the GF exists. The DE for G(x, y) is G" = 0 for x -::f. y, whose solutions are G(x, y)

= {CIX + C2 if a dtx+d2

:::::: x < y,

(20.35)

ify <.x s,b.

Continuityatx = y givesqy +c2 = dlY +dz or bj y +bz discontinuity of d G/ dx at x = y gives

1 dt - Cj = - - = I P2w

= Owithbj

= Ci -di- The

bl =-1

=}

assuming that w = 1. From the equations above we also get bl = y. G(x, y) must also satisfy the given BCs. Thus, G(a, y) = 0 = G(b, y). Since a S, y and b ::: y, we obtain CI a + C2 = 0 and dl b + dz = 0, or,after substituting Ci = hi + di, ad l +d2 = a - y,

bd, +d2 = O.

The solution to these equations is dl = (y - a)/(b - a) and d2 With bl. hz. di. anddz as given above,we find b-y Cj=bl+dl=--b-a

and

=

-b(y - a)/(b - a).

b-y b-a -·

C2 =b2+ d2 =a-

Writiog Equation (20.35) as G(x, y) = (Cjx

+ C2)O(y -

x)

+ (djx + d2)O(X -

y)

and using the identity O(y - x) = 1 - O(x - y), we get G(x, y) = Cjx + C2 - (blx

+ ~)O(x -

y).

Using thevaluesfoundfor the b's andc's, we obtain

Y) + (x -

G(x, y) = (a - x) (b b-a

y)O(x - y),

which is the same as the GF obtained in Example 20.1.4.

III

20.3.6. Example. Letus findthe GFfor Lx = d 2 /dx 2+1 withtheBCsu(O) = u(,,/2) = O. The general solution ofLx[u] = 0 is u(x)

= A sin x + Bcosx.

!fthe BCs are imposed, we get u = O. Thus, G(x, y) exists. The generalfonn ofG(x, y) is G(x, y) = {CI sinx

+ C2 cosx

dl sinx +d2COSX

if 0::::; x < y, if Y < x::::; rr/2.

(20.36)

572

20. GREEN'S FUNCTIONS IN ONE OIMENSION

Continuity of G at x = y gives hI sin y + b: cos y = a with hi = Ci - dj. The discontinuity of the derivative of G at x = y gives hI cos y - b: sin Y = -I, where we haveset w(x) = 1. Solving these equations yields hI = - cos y and ~ = siny. The Bes give G(O,y)=O ~ c2=0 ~ dz=-b2=-siny,

G(rr/2,y)=0

~

d,=O ~ q=-b,=-cosy.

Substituting in Equatioo(20.36) gives - COS Y Sin X

G(x,y) = { . -smycosx

if x -c y, if y .c x,

or, usingthe thetafunction, G(x, y) = -O(y - x) cos y sin x - O(y - x) siny cosx

= -[I - O(x - y)] cos Y sin x - O(x - y) siny cos x = -cosy sin x

+ O(x -

y) sin(x - y).

It is instructiveto verify directly that G(x, y) satisfies Lx[G] = Lx[G] = -cosy

(::z +I)

(~+,) sin x +

~(x

- y):

[O(x - y) sin(x - y)]

~O

dZ

= dx 2 [O(x - y) sin(x - y)] =

~[~(x dx

+ O(x -

y) sin(x - y)

y) sin(x - y) +O(x - y) cos(x - y)] •

+ O(x -

y) sin(x - y).

~O

The first term vanishes because the sine vanishes atthe only pointwherethedeltafunction is nonzero. Thus,we have Lx[G] = [~(x - y) cos(x - y) - O(x - y) sin(x - y)]

=8(x -

+ O(x -

y) sin(x - y)

y)

because the delta function demands that x = y, for which cos(x - y) = I. The existence and nniqueness of the Green's function G(x, y) in conjnnction with its properties and its adjoint, imply the existence and uniqueness of the adjoint Green's function g(x, y). Using this fact, we can show that the condition for the absence of a nontrivial solution for Lxlu] = 0 is also a necessary condition for the existence of G(x, y). That is, if G(x, y) exists, then Lxlu] = 0 implies that u = O. Suppose G(x, y) exists; then g(x, y) also exists. In Green's identity let v = g(x, y). This gives an identity:

l

b

w(x)g*(x, y)(Lxlu]) dx =

=

l l

a

b

w(x)u(x)(Lllg])* dx b

w(x)u(x)

8(x - y) w(x)

dx = u(y).

20.3 GREEN'S FUNCTIONS FOR SOlDOS

In particular, if Lx[u] result.

573

= 0, theu u(y) = 0 for all y. We have proved the following

20.3.7. Proposition. The DE Lx[u] = 0 implies that u sa 0 corresponding to lx and the homogeneous Bes exist.

if and only if the GF

lt is sometimes stated that the Green's function of a SOLDO with constant coefficients depends on the difference x - y. This statement is motivated by the observation that if u(x) is a solution of d 2u

Lx[u]

du

= a2 dx 2 + al dx + aou = f ix),

then u(x - y) is the solution of a2u" + alu' + aou = I(x - y) if ao, ai, and a2 are constant. Thus, if G(x) is a solution of Lx[G] = 8(x) [again assuming that w(x) I], then it seems that the solution of Lx[G] 8(x - y) is simply G(x - y). This is clearly wrong, as Examples 20.3.5 and 20.3.6 showed. The reason is, of course, the BCs. The fact that G (x - y) satisfies the right DE does not guarantee that it also satisfies the right BCs. The following example, however, shows that the conjecture is true for a homogeneous initial value problem.

=

=

20.3.8. Example. The mosl general fonn for the OF is CI UI (X) +C2U2(x) G(x,y) = { dlul(x) +d2U2(x)

if a ~ x < Y, if Y < x ~ b.

The IVP condition G(a, y) = 0 = G'(a, y) implies CIUI(a) +c2u2(a) = 0

and

Linear independence of Ul andU2 implies UI (a)

det ( u;(a) Hence, C1 =

U2(a») ui(a) = W(a; uI, U2) "" O.

C2 =

G(x, y) =

0 is theonly solution. Thisgives

o {dIUI(X) +d2U2(x)

ffa5x
if y < x ~ b.

(20.37)

Continuity of G at x = y yields dlUI (y) + d2U2(y) = 0, while the discontinuity jump conditionin the derivative gives diU; (y) + d2ui(Y) = I. Solvingthese two equations, we get dl =

U2(Y) , U;(Y)U2(Y) - ui(y)uI(Y)

d2=

U; (Y)U2(Y) - ui(y)uI (y)'

Substituting this in (20.37)gives G(x, y) = [U2(Y)UI (x) - UI (Y)U2(X)] e(x _ y). u; (Y)U2(Y) - ui(y)uI (y)

(20.38)

574

20. GREEN'S FUNCTIONS IN ONE OIMENSION

Equation (20.38) holdsfor any SOLDO with the givenBes. Wenowuse the fact that the SOLDO has constantcoefficients. In thatcase, we knowtheexactformof zej and "2. There are two cases to consider: 1. If the characteristic polynomial of Lx has two distinct roots Al and}..2, then Ul (x) = e AtX and U2(X) = eA2X . Writing Al = a + b and)..2 = a - b andsubstituting the exponential functions andtheirderivatives in Equation (20.38) yields e Ca - b)Ye ca + b)X _ e(a+b)ye(a-bh]

G(x, y) = [

2be2ay

= 2~ [e(a+b)(X-y) -

e(x - y)

e(a-b)(x-Y)j e(x - y),

which is a function of x - y alone. 2. If Al = ).2 = A, then Ul(X) = eAx , u2(x) = xe Ax , and substitution of these functions in Equation (20.38) gives G(x, y)

= (x -

y)e'(x-Y)e(x - y).

III

20.3.3 Inhomogeneous Bes

GF solves inhomogeneous BCs as well

So far we have concentrated on problems with homogeneous BCs, Ri [u] = 0, for i = 1,2. What if the BCs are inhomogeneous? It turns out that the Green's function method, even though it was derived for homogeneous BCs, solves this kind of problem as well! The secret of this success is the generalized Green's identity. Suppose we are interested in solving the DE Lx[u] = f(x)

with

Ri[U] = Yi

for i = 1,2,

and we have the GF for Lx (with homogeneous BCs, of course). We can substitute v = g(x, y) = G*(y, x) in the generalized Green's identity and use the DE to

obtain

l

b

w(x)G(y, x)f(x) dx

-l

b

w(X)u(x)(Lr[gj)* dx = Q[u, g*(x,

Y)]I;~~,

or, using d[g(x, y)] = 8(x - y)/w(y), u(y) =

t

w(x)G(y, x)f(x) dx - Q[u, g*(x,

Y)]I;~~.

To evaluate the surface term, let us write the BCs iu matrix form [see Equation (20.17)]: AUa AGa

+ BUb = + BGb =

Y 0

'* = B-Iy - B-IAua , '* A'(B')-IOb9); + Oa9; = 0, Ub

(

20.3 GREEN'S FUNCTIONS FOR SOlOOS

575

where Y is a column vector composed of YI and Y2, and we have assumed that G(x, y) and g*(x, y) satisfy, respectively, the homogeneons BCs (with Y = 0) and their adjoints. We have also assumed that the 2 x 4 matrix of coefficients has rank 2, and withont loss of generality, let Bbe the invertible 2 x 2 submatrix. Then, assuming the general form of the surface term as in Equation (20.15), we obtain * toaga* Q[u,g *(x,Y )][x~b x=a=ubtob9b-ua

= ( B-I Y - B-I A Ua) ' ab9b* - uta Qa 9 a* = y'(B,)-I Ob9); - U~,[At(B')-IOb9); +Oa9~] = 0 because g* satisfies homogeneous adjoint Be

(20.39) . where

*

gb sa

(

g*(b, y) ) ( G(y, b) ) a * = ax g (x, y)[x=b axG(y,X)[x~b

a

.

It follows that Q[u, g*(x, Y)]I;'~~ is given entirely in terms of G, its derivative, the coefficient functions of the DE (hidden in the matrix 0), the homogeneous BCs (hidden in B), and the constants YI and Y2. The fact that g* and agO lax appear to be evaluated at x = b is due to the simplifying (but harmless) assumption that B is invertible, i.e., that u(b) and u'(b) can be written in terms of ural and u'(a). Of course, this may not be possible; then we have to find another pair of the four quantities in terms of the other two, in which case the matrices and the vectors will change but the argument, as well as the conclusion, will remain valid. We can now write u(y) =

l

b w(x)G(y,x)j(x)dx _ ytM9*,

(20.40)

where a general matrix Mhas been introduced, and the subscript b has beenremoved to encompass cases where submatrices other than Bare invertible. Equation (20.40) shows that u can be determined completely once we know G(x, y), even though the BCs are inhomogeneous. In practice, there is no need to caicolate M. We can use the expression for Q [u, gO] obtained from the Lagrange identity of Chapter 13 and evaluate it at b and a. This, in general, involves evaluating u and G and their derivatives at a and b. We know how to handle the evaluation of G because we can actually constroct it (if it exists). We next find two of the four' quantities corresponding to u in terms of the other two and insert the result in the expression for Q[u, gO]. Equation (20.39) then guarantees that the coefficients of the other two terms will be zero. Thus, we can simply drop all the terms in Q[u, gO] containing a factor of the other two terms.

(

576

20. GREEN'S FUNCTIONS IN ONE OIMENSION

Specifically, we use the conjunct for a formally self-adjoint SOLDO [see Equation (20.26)] and g*(x, y) = G(y, x) to obtain

u(y) =

l

b w(x)G(y,x)f(x)dx

eo

du ]}X~b - { p(x)w(x) [G(y, x) dx - u(x)h(y, x) x~a Intercbanging x and y gives

u(x) =

l

b

w(y)G(x, y)f(y) dy (20.41)

eo(x, y) + {p(y)w(y) [u(y)-a y

du]}Y=b G(x, y)-d

y

y=a

This equation is valid only for a self-adjoint SOLDO. That is, using it requires casting the SOLDO into a self-adjoint form (a process that is always possible, in light of Theorem 13.5.4). By setting f(x) = 0, we can also obtain the solution to a homogeneous DE Lx [u] = 0 that satisfies the inhomogeneous Bes. 20.3.9. Example. Let us find the solution of the simple DE d 2u/dx2 = f(x) subject to the simple inhomogeneous BCs u(a) = Yl and u(b) = Y2. The GF for this problem has been calculated in Examples 20.1.5 and 20.3.5.Let us begin by calculating the surfaceterm in Equation (20.41). We have p(y) = I, and we set w(y) = I, then surface term = u(b) oG I - G(x, b)u'(b) - u(a) oG I oy y~a oy y=b

= Y2

oG I oy y~b

-

Yl oG I oy y~a

+ G(x, a)u'(a) -

+ G(x, a)u'(a)

G(x, b)u'(b).

That theunwanted (andunspecified) terms are zerocanbe seenby observing that G(x, a) = g*(a, x) = (g(a, x))*, and that g(x, y) satisfies the BCsadjoint to the homogeneous BCs (obtained when Yt = 0). In this particular and simple case, the BCs happen to be selfadjoint (Dirichlet BCs). Thus, u(a) = u(b) = 0 implies that g(a, x) = g(b, x) = 0 for all x E [a, b]. (In a more general case thecoefficient ofu'(a) wouldbe morecomplicated, but still zero.) Thus, we fioally have \ surface term = Y2 oG I oy y=b

- Yl oG I

oy y=a

.

Now,using the expression for G(x, y) obtained in Examples 20.1.5 and 20.3.5, we get 8G

a-x

-oy = -b-a -- -

9(x - y) - (x - y)8(x - y) "

x-a

=-b-a

=0

Thus,

so I

BY y=b =

x-a b-a'

OGI

x -a b -- - 1 -x-- oy y=a - b-a - b-a'

8(x - y).

20.4 EIGENFUNCTION EXPANSION OF GREEN'S FUNCTIONS

577

Substituting in Equation (20.41), we get

f

u(x) =

G(x,y)f(y)dy +

~=~IX+ bY~=:Y2.

(Compare this with the result obtained in Example 20.1.5.)

III

Green's functions have a very simple and enlightening physical interpretation. An inhomogeneous DE such as Lx[u] = f(x) can be interpreted as a black box (L,) that determines a physical quantity (u) when there is a source (f) of that physical quantity. For instance, electrostatic potential is a physical quantity whose source is charge; a magnetic field has an electric current as its source; displacements and velocities have forces as their sources; and so forth. Applying this interpretation and assuming that w(x) = 1, we have G(x, y) as the physical quantity, evaluated at x when its source 8 (x - y) is located at y. To be more precise, let us say that the strength of the source is 81 and it is located at Yl; then the source becomes 818(x - Yl). The physical quantity, the Green's function, is then 81G(X, Yl), because of the linearity of Lx: If G(x, y) is a solution of Lx [u] = 8 (x - y), then 81G(x, Yl) is a solution of Lx[u] = 818(x - y). If there are many sources located at {yil[:,1 with corresponding strengths {8il[:,1' then the overall source f as a function of x becomes f(x) = L,[:,t 8i8(X - Yi), and the corresponding physical quantity u(x) becomes u(x) = L,[:,1 8iG(X, Yi). Since the source S; is located at Yi, it is more natural to define a function 8 (x) and write S; = 8 (Yi). When the number of point sources goes to infinity and Yi becomes a smooth continuous variable, the sums become integrals, and- we have f(x) =

l

b

u(x) =

8(y)8(x - y) dy,

l

b

8(y)G(x, y) dy.

The first integral shows that 8 (x) = f (x). Thus, the second integral becomes f(y)G(x, y) dy which is precisely what we obtained formally.

u(x) =

20.4

J:

Eigenfunction Expansion of Green's Functions

Green'sfunctions areinverses of differential operators. Inverses of operators in a Hilbert space are best studied in terms of resolvents. This is because if an operator A has an inverse, zero is in its resolvent set, and

Thus, it is instructive to discuss Green's functions in the context of the resolvent of a differential operator. We will cousider only the case where the eigenvalues are discrete, for example, when Lx is a Sturm-Liouville operator. Formally, we have (L - Al)R)JL) = 1, wbich leads to the DE (Lx - A)R).(x, y) =

8(x - y) w(x)

,

578

20. GREEN'S FUNCTIONS IN ONE DIMENSION

where R).(x, y) = (xl R,,(L) ly). The DE simply says that R,,(x, y) is the Green's function for the operator Lx - A. So we can rewrite the equation as A)G,,(X, y) = 8(x - y)jw(x) where Lx - A is a DO having some homogeneous BCs. The GF G" (x, y) exists if and only if (Lx - A)[u] = 0 has no nontrivial solution, which is true only if A is not an eigenvalue of Lx. We choose the BCs in such a way that Lx becomes self-adjoint. Let {Anl~l be the eigenvalues of the system Lx[u] = AU, {Ri[U] = OIT=l' and let the u~k) (x) be the corresponding eigenfunctions. The index k distinguishes among the linearly independent vectors corresponding to the same eigenvalue An. Assuming that L has compact resolvent (e.g., a Sturm-Liouville operator), these eigenfunctions form a complete set for the subspace of the Hilbert space that consists of those functions that satisfy the same BCs as the u~k) (x). In particnlar, G,,(x, y) can be expanded in terms of u~k)(x). The expansion coefficients are, of course, functions of y. Thus, we can write

a, -

00

G,,(x, y) =

L ~>~k)(y)u~k)(x) k n=l

J:

where a~k)(y) = w(x)u~(k)(x)G,,(x, y) dx. Using Green's identity, Equation (20.30), and the fact that An is real, we have

Ana~k)(y) =

1 W(X)[AnU~k\X)]'G,,(x, l y){Lx[u~k)(x)]}· l w(x)[u~k)(x)]'Lx[G,,(x, l w(x)u~(k)(x) + u~(k)(y) + l w(x)u~(k)(x)G).(x, b

.

y)dx

b

=

w(x)G).(x,

dx

b

=

y)] dx

b

=

[8(X - y) w(x)

a

AG,,(X, y)] dx

b

=

A

y) dx

= u~(k)(y) + Aa~k)(y). Thus, a~k)(y) = u~(k)(y)j(An - A), and the expansion for the Green's function is _ ' " ~ u~(k)(y)u~k\x) G" (x, y ) - L... L... A _ A . k n=l

(20.42)

n

This expansion is valid as long as An i' A for any n = 0, I, 2, .... But this is precisely the condition that ensures the existence of an inverse for L - A1. An interesting result is obtained from Equation (20.42) if A is considered a complex variable. In that case, G" (x, y) has (infinitely many) simple poles at

20.4 EIGENFUNCTION EXPANSION OF GREEN'S FUNCTIONS

579

c;

{A"}~l' The residue at the pole A" is - Lk u:(k)(y)u~k)(x). If is a contour having the poles {A"};:'=l in its interior, then, by the residue theorem, we have

In particular, if we let m -+

00,

we obtain S(x - y)

w(x)

,

(20.43)

where Coo is any contour that encircles all the eigenvalues, and in the last step we used the completeness of the eigenfunctions. Equation (20.43) is the infinitedimensional analogue of Equation (16.12) with f (A) = 1 when the latter equation is sandwiched between (x I and Iy). 20.4.1. Example. Consider the DO Lx = d 2/dx2 with BCs u(O) = ural =

o. This is

an S-L operator with eigenvalues and normalized eigenfunctions _ (mr)2 An-

a

(2 .

Un(X)=v~sm

and

aX

(mr )

for n = 1,2, ....

Equation (20.42) becomes G x

)J , y)

__ ~ ~ sin(mrx/a) sin(mry/a) L. ' ( /)2 a n=l n1f a ,

-

I\, -

which leads to

-1 2rri

i

Coo

G, (x )Yd A = - 1 '.

i [-- ~

2:rri Coo

= __ 1

2:Jr:i = _

a

I:

sin (mr

n=l

a

x) sin (""a y) ie 1.

oo

f: (n" x) (n" y) (~) f: sinC" x) SinC" y).

(~) a

= -

(~)

2 L. sin(mrx/a)sin(mry/a)] d A a n=l A - (mrja)2

sin

sin

Res [

n=l

a

a

a n=l

a

a

dA A - (nJrja)2 1

'). - (mf/a)

2] l=An

The RHS is recognized as -8(x - y).

eigenfunclion expansion of GF

If zero is not an eigenvalue of Lx, Equation (20.42) yields 00

u:(k) (y )u~k) (x)

G(x,y)=Go(x,y)=LL k n=l

'

(20.44)

An

which is an expression for the Green's function of Lx in terms of its eigenvalues and eigenfunctions.

580

20. GREEN'S FUNCTIONS IN ONE OIMENSION

20.5

Problems

Lxu(x) = dufdx = f(x) subject to = a. Hint: Consider the function vex) = u(x) - a. 20.2. Solve the problem of Example 20.1.4 subjectto the BCs u(a) = u'(a) = O.

20.1. Using the GF method, solve the DE the BC u(O)

Show that the corresponding GF also satisfies these BCs.

20.3. Show that the IVP with data 10; 0, 0, ... , OJ has only u '" 0 as a solution. Hint: Assume otherwise, add u to the solution of the inhomogeneous equation, and invoke uniqueness. 20.4. In this problem, we generalize the concepts of exactness and integrating factor to a NOLDE. The DO Lin) es I:~=o pkCx)dk /dx k is said to be exact if there exists a DO Min-I) sa I:~;;6 ak(x)dk /dx k such that VUE

en[a, b].

(a) Show that Lin) is exact iff I:~~o( _I)mdmPm/dxm = O. (b) Show that there exists an integrating factor for Lin)-that is, a function /L(x) such that /L(x)Lin) is exact-if and only if /L(x) satisfies the DE

The DO Nin) is the formal adjoint of L¥').

20.5. Assuming that

Lx[u] = 0 has no nonttivial solution, show that the matrix

where UI and U2 are independent solutions of Lx[u] = 0 and Ri are the boundary functionals, has a nonzero determinant. Hint: Assume otherwise and show that the system of homogeneous linear equations aRI[uI] + ,BRI[U2] = 0 and aR2[u!l + ,BR2[U2] = 0 has a nontrivial solution for (a,,8). Reach a contradiction by considering u = aUI + ,Bu2 as a solution of Lx[u] = O.

20.6. Determine the formal adjoint of each of the operators in (a) through (d) below (i) as a differential operator, and (ii) as an operator, that is, including the BCs. Which operators are formally self-adjoint? Which operators are self-adjoint? (a) Lx = d 2/dx2 + I in [0, 1] with BC, u(O) = u(l) = O. (b) Lx = d 2/dx2 in [0, 1] withBCs u(O) = u'(O) = O. (c) Lx = dfdx in [0, 00] withBCs u(O) = O. (d) Lx = d 3/dx3 - sinxd/dx + 3 in [0, IT] with BCs u(O) = u'(O) = 0, u"(O) 4u(IT)

= O.

20.5 PROBLEMS

581

20.7. Show that the Dirichlet, Neumann, general umnixed, and periodic BCs make the following formally self-adjoint SOLDO self-adjoint: Lx =.!..~ wdx

(p~) +q. dx

20.8. Using a procedure similar to that described in the text for SOLDOs, show that for the FOLDO Lx = PI d j dx (a) the indefinite GF is G(x, y) sa

where

+ PO

{'(y) [e(x - y)] PI(Y)w(y) {'(x) {'(x) = exp

[I

X

+ C(y),

po(t)

]

- - dt , PI(t)

(b) and the GF itself is discontinuous at x = y with

. Inn [G(y

,-+0

+ E, y) -

G(y - E, y)] =

1 () ()' PI Y w Y

(c) For the homogeneous BC R[u] sa O
+ 0<2u'(a) + f3tu(b) + fhu'(b) =

0

construct G(x, y) and show that G(x, y) =

I () () ( ) v(x)e (x - y) PI Y w Y v Y

+ C(y)v(x),

where v(x) is any solution to the homogeneous DE Lx[u] = 0 and C(y) =

+

f!(v(b) {hv'(b) , R[v]PI (y)w(y)v(y)

(d) Show directly that Lx[G] =

~(x

with R[v]

01 0

- y)jw(x).

20.9. Let Lx be a NOLDO with constant coefficients. Show that if u(x) satisfies Lx[u]

= f(x), then u(x -

y) satisfies Lx[u]

= f(x

- y). (Note that no BCs are

specified.)

20.10. Find the GF for Lx

= d 2jdx 2 + 1 with BCs u(O) = u'(O) = O. Show that

it can be written as a function of x - y only.

20.11. Find the GF for Lx = d 2jdx 2 + k 2 with BCs u(O) = u(a) 20.12. Find the GF for Lx = d 2jdx 2 - k 2 with BCs u(oo)

= O.

= u(-oo) = O.

20.13. Find the GF for Lx = (djdx)(xdjdx) given the condition that G(x, y) is finite at x = 0 and vanishes at x = I.

582

20. GREEN'S FUNCTIONS IN ONE DIMENSION

20.14. Evaluate the GF and the solutions for each of the following DEs in the interval [0, I]. (a) u" - k 2u = f; u(O) - u'(O) = a, u(l) = b. (b) u" = f; u(O) = u'(O) = O. (c) u" + 6u' + 9u = 0; u(O) = 0, u'(O) = 1. (d) u" + !J)2u = f(x), for x > 0; u(O) = a, u'(O) = b. (e) u(4) = f; u(O) = 0, u'(O) = 2u'(I), u(l) = a, u"(O)

= O.

20.15. Use eigenfunction expansion ofthe GF to solve the BVP u"

= x, u (0) = 0,

u(l) - 2u'(I) = O.

Additional Reading I. Dennery, P. and Krzywicki, A. Mathematics for Physicists, Harper and Row, 1967. Contains an exposition of Green's functions for second-order linear differential equations. 2. Roach, G. Green's Functions, Van Nostrand, 1970. A readable introduction to Green's functions, especially those we have called "indefinite" (no bonndary conditions specified).

21

_

Multidimensional Green's Functions: Formalism

The extensive study of Green's functions in one dimension in the last chapter has no doubt exhibited the power and elegance of their use in solving inhomogeneous differential equations. If the differential equation has a (unique) solution, the GF exists and contains all the information necessary to build it up. The solution results from operating on the inhomogeneous term with an integral operator whose kernel is the appropriate Green'sfunction. The Green's function's very existence depends on the type ofBCs imposed. We encountered two types of problems in solving ODEs. The first, called initial value problems (NPs), involves fixing (for an nth-order DE) the value of the solution and its first n - I derivatives at a fixed point. Then the ODE, if it is sufficiently well-behaved, will determine the values of the solution in the neighborhood of the fixedpoint in a unique way. Becauseof thisuniqueness, Green's functions always exist for IVPs. The second type of problems, called boundary value problems (BVPs), consists-when the DE is second order-of determining a relation between the solution and its derivative evaluated at the boundaries of some interval [a, b]. These boundary values are relations that we denoted by Ri[U] = Yi, where i = 1,2. In this case, the existence and uniqueness of the Green's function are not guaranteed. There is a fundamental (topological) difference between a boundary in one dimension and a boundary in two and more dimensions. In one dimension a boundary consists of only two points; in 2 and higher dimensions a boundary has infinitely many points. The boundary of a region in jR2 is a closed curve, in jR3 it is a closed surface, and in jRm it is called a hypersurface. This fundamental difference makes the study of Green's functions in higher dimensions more complicated, but also richer and more interesting.

584

21. MULTIDIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

21.1 Properties of Partial Differential Equations This section presents certain facts and properties of POEs, in particular, how BCs affect their solutions. We shall discover the important difference between OOEs andPOEs:

21.1.1. Box. The existence of a solution to a PDE satisfying a given depends on the type ofthe PDE.

Be

We shall be concerned exclusively with a linear POE. A linear POE of order

M in m variables is of the form Lxlu] = f(x)

where

(21.1)

where the following notation has been used:

x=

(Xl, ... , X m ) ,

IJI = h

+ Iz + ... + i«.

J=(h, .. ·,jm), alJI a lJI =

ax!

axil ax42 ••. ax!:' '

the ik are nonnegative integers; M, the order of the highest derivative, is called the order of the POE. The outer sum in Equation (21.1) is over IJI; once IJI is fixed,

the inner summation goes over individual has to equal the given IJ I. principal part of a POE

A'8 with the restriction that their sum

The principal part of Lxis (21.2)

Cauchy data and Cauchy problem

The coefficients a j and the inhomogeneous (or source) term f are assumed to be contiouous functions of their arguments. We consider Equation (21.1) as an NP with appropriate initial data. The most direct generalization of the IVP ofordinary differential equation theory is to specify the values of u and all its normal derivatives of order less than or equal to M - 1 on a hypersurface r of dimension m - 1. This type of initial data is called Cauchy data, and the resulting IVP is known as the Cauchy problem for Lx. The reason that the tangential derivatives do not come into play here is that once we know the values of u on r, we can evaluate u on two neighboring points on I', take the limit as the points get closer and closer, and evaluate the tangential derivatives.

21.1 PROPERTIES OF PARTIAL DIFFERENTIAL EQUATIONS

21.1.1

585

Characteristic Hypersurfaces

In contrast to the IVP in one dimension, the Cauchy problem for arbitrary Caucby

data may not have a solution, or if it does, the solutionmay not be unique.

21.1.2. Box. The existence and uniqueness of the solution of the Cauchy problem depend crucially on the hypersurface r and on the type ofPDE.

tangential coordinates

We assume that r can be parametrized by a set of m functions of m - 1 parameters. These parameters can be thought of as generalized coordinates of points of r. Consider a point P on r. Introduce m - 1 coordinates ~2, ... , ~m, called tangential coordinates, to label points on r. Choose, by translation if necessary, coordinates in such a way that P is the origin, with coordinates (0, 0, ... ,0). Now let v = ~l stand for the remaining coordinate normal to r. Usually ~i is taken to be the lth coordinate of the projection of the point on r onto the hyperplane tangent to F' at P. As long as we do not move too far away from P, the Cauchy data on r can be written as au aM-lu u(O, ~2, ... , ~m), av (0, ~2, ... , ~m), ... , av M - 1 (0, ~2, ... , ~m).

Using the chain rule, au/ax, = L:7~1 (au/a~j )(a~j /aXi), where ~l = v, we can also determine the first M -·1 derivatives of u with respect to Xi. The fundamental question is whether we can determine u uniquely using the above Cauchy data and the DE. To motivate the answer, let's look at the analogous problem in one dimension. Consider the Mth-order linear ODE dMu

Lx[u]

du

= aM(x) dx M + ... + al(x) dx + ao(x)u = f(x)

(21.3)

with the following initial data at xo: {u(xo), u'(xo), ... , u(M-I)(xO)}. If the coefficients {ak(x)}~o and the inhomogeneous term f(x) are continuous and if au (xo) i' 0, then Theorem 20.2.3 implies that there exists a unique solution to the IVP in a neighborhood of xo. For aM (xo) i' 0, Equation (21.3), the initial data, and a knowledge of f(xo) give U(M)(xo) uniquely. Having found u(M)(xO), we can calculate, with arbitrary accuracy (by choosing Ax small enough), the following set of new initial data at Xl = Xo

+ Ax:

U(XI) = u(xo)

+ u'(xo)AX,

... , u(M-I)(XI) = u(M-I)(xO)

+ U(M)(xo)Ax.

Using these new initial data and Theorem 20.2.3, we are assured of a unique solution atxl. Since au (x) is assumed to be continuous for Xl, for sufficiently small

586

21. MULTIDIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

Sx, aM(xo) is nonzero, and it is possible to find newer initial data at X2 = Xl + L\.x. The process can continue until we reach a singularity ofthe DE, a point where aM (x) vanishes. We can thus construct the unique solution of theIVP in an interval (xo, b) as long as aM(x) does not vanish anywhere in [xo, b]. This procedure is analogous to the one used in the analytic continuation of a complex function. For aM(xO) = 0, however, we cannot calculate u(M)(xO) unambiguously. In such a case the LHS of (21.3) is completely determined from the initial data. If the LHS happens to be equal to f(xo), then the equation is satisfied for any U(M) (xo), i.e., there exist infinitely many solutions for U(M) (xo); if the LHS is not equal to f (xo), there are no solutions. The difficulty can be stated in another way, which is useful for generalization to the m-dimensional case:

21.1.3. Box. If au (xo) = 0 in (21.3), then the initial data determine the function Lx[uj.

Let us now return to the question of constructing u and investigate conditions under which the Cauchy problem may have a solution. We follow the same steps as for the IVP for ODEs. To construct the solution numerically for points near P but away from r (since the function is completely determined on r, not ouly its Mth derivative but derivatives of all orders are known on I"), we must be able to calculate aMujav M at P. This is not possible if the coefficient of aMujav M in Lx[u] is zero when Xl, ••• , X m is written in terms of v, ~2, '" ,~m. Whenthis happens, Lx[u j itself will be determined by the Cauchy data. This motivates the following definition. characteristic hypersurface

21.1.4. Definition. IfLx[uj can be evaluated at a point P on r from the Cauchy data alone, then r is said to be churacteristic for Lx at P. Ifr is characteristicfor all its points, then it is called a characteristic hypersurface for Lx. The Cauchy problem does not have a solution at a point on the characteristic hypersurface. The following theorem characterizes I': 21.1.5. Theorem. Let r be a smooth (m - I)-dimensional hypersurface. Let Lx[u] = f be an Mth-order linear PDE in m variables. Then r is characteristic at PEr if and only if the coefficient of aMujav M vanishes when Lx is expressed in terms ofthe normal-tangential coordinate system (v, ~2, ... , ~m). One can rephrase the foregoing theorem as follows: 21.1.6. Box. The hypersurface r is not characteristic at P if and only if all M th-order partial derivatives ofu with respect to [xili=l are unambiguously determined at P by the DE and the Cauchy data on r.

21.1 PROPERTIES OF PARTIAL OIFFERENTIAL EQUATIONS

Characteristic "hypersurtaces" of ODEs are points!

587

In the one-dimensional case the difficulty arose when aM(xO) = O. In the language being used here, we could call xo a "characteristic point." This makes sense because in this special case (m = I), the hypersurfaces can ouly be of dimension O.Thus, we can say that in the neighborhood ofa characteristic point, the IVP has no well-defined solution. I For the general case (m > I), we can similarly say that the Cauchy problem bas no well-defined solution in the neighborhood of P if P happens to lie on a characteristic hypersurface of the differential operator. Thus, it is important to determine the characteristic hypersurfaces of PDEs. 21.1.7. Example. Lei us consider the first-orderPDE in two variables

au + hex, y)au + F(x, y, u) =

Lx[u] = a(x, y)ax

ay

0

(21.4)

where F (x, y, u) == c(x, y)u + d (x, y). For this discussion the form of F is irrelevant. Wewishto find thecharacteristic hypersurfaces (in thiscase,curves) of L. TheCauchy data consist of a simpledetermination of u on r. By Theorem 21.1.5, we need to derive relations that ensure that au/ax and 8uj8y cannot be unambiguously determined at P = (x, y). Using an obviousnotation, the PDE of Equation (21.4) gives au -F(P, u(P)) = a(P)-(P) ax

au

+ b(P)-a (P). y

+ dx, y + dy) lies on the curve F, then au au u(P) = dx-(P) +dy~(P).

On the oiher band, if Q es (x u(Q) -

ax

ay

The Cauchy data determine the LHS of both of the preceding equations. Treating these equations as a system of two linear equations in two nnknowns, au/ax(p) and au/ay(p), we conclude that the systemhas a unique solution if andonly if the matrix of coefficients is invertible. Thus, by Box 21.1.6, I' is a characteristic curveif andonly if dx

det ( alP)

dy ) = b(P) dx - alP) dy = 0, b(P)

or dy/dx = hex, y)/a(x, y), assaming ihat a(x, y) # O. Solving this FaDE yields y as a function of x, thus determining thecharacteristic curve. Note that a general solution of this III FODEinvolves anarbitrary constant, resulting in a familyof characteristic curves.

SofiaVasilyevnaKovalevskaya (1850--1891) is consideredthe greatest womanmaihematiciao prior to the twentieth century. She grewup in a well-educated family of the Russian nobility, herfather beinganartillery general andreputed to be a descendant of a Hungarian king, Mathias Kervin. Sonja was educated by a British governess andenjoyed life at the large country estate of herfather's family, although the rather progressive thinking of the Kovalevsky sisters didnot always meet withapproval fromtheir father. 1Herelies thecrucial difference betweenODEsandPDEs: All ODEshavea universal characteristic hypersurface, i.e., apoint. PDEs,on theotherhand, canhavea variety of hypersurfaces.

588

21. MULTIOIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

Sonja has written of two factors that attracted her to the study of mathematics. The first was her Uncle Pyotr, who had studied the subject on his own and would speak of squaring the circle and of the asymptote, as well as of many other things that excitedher imagination. The second was a curious "wallpaper" that was used to cover one of the children's rooms at Polibino, which turned out to be lecture notes on differential and integral calculus that had been purchased by her father in student days. These sheets fascinated her and she would

spendhours trying to decipherseparatephrases and to find the properorderingof the pages. In the autumn of 1867 Sonja went to St. Petersburg, where she studied calculus with Alexander Strannolyubsky, a teacher of mathematics at the naval school. While there, she consulted the prominent Russian mathematician Chebyshevabout her mathematical studies, but since Russian universities were closed to women, there seemed to be no way that she could pursue advanced studies in her native land. In order to escape the oppression of women common in Russia at the time, young ladies of ambition and ability would often arrange a marriage of convenience in order to allow study at a foreign university. At the age of 18, Sonya arranged such a marriage with Vladimir Kovalevsky, a paleontologist, and in 1869 the couple moved to Heidelberg, where Sonja took courses from Kirchhoff, Helmholtz, and others. Two years later she went to Berlin, where she worked with Weierstrass,who tutored her privately, since she, as a woman, was not allowed to attend lectures. The three papers she published in the next three years earned her a doctorate in absentia from the University of G5ttingen. Unfortunately, even that distinction was not sufficient to gain her a university position anywhere in Europe, despite strong recommendation from the renowned Weierstrass. Her rejections resulted in a six-year period during which time she neither undertook research nor replied to Weierstrass's letters. She was bitter to discover that the best job she was offered was teaching arithmetic to elementary classes of schoolgirls, and remarked, "I was unfortunately weak in the multiplication table." The existence and uniqueness of solutions to partial differential equations occupied the attention of many notable mathematicians of the last century, including Cauchy, who transformed the problem into his method of majorant functions. This method was later extended and refined by Kovalevskaya to include more general cases. The result was the now-famous Cauchy-Kovalevskaya theorem. She also contributed to the advancement of the study of Abelian integrals and functions and applied her knowledgeof these topics to problems in physics, including her paper "On the Rotation of a Solid Body About a Fixed Point," for which she won a 5000-franc prize. She also performed some investigations into the dynamics of Saturn's rings, inspiring a sonnet in which she is named "Muse of the Heavens." In 1878, Kovalevskaya gave birth to a daughter, but from 1880 increasingly returned to her study of mathematics. h11882 she began work on the refraction of light, and wrote three articles on the topic. In the spring of 1883, Vladimir, from whom Sonja had been separated for two years, committed suicide. After the initial shock, Kovalevskaya immersed herself in mathematical work in an attempt to rid herself of feelings of guilt. MiUag-Leffler managed to overcome opposition to Kovalevskaya in Stockholm, and obtained for her a position as privat docent. She began to lecture there in early 1884, was appointed to a fiveyear extraordinary professorship in June of that year, and in June 1889 became the third woman ever to hold a chair at a European university.

21.1 PROPERTIES OF PARTIAL OIFFERENTIAL EQUATIONS

589

During Kovalevskaya's years at Stockholm she carried out important research, taught courses on the latest topics in analysis, and became an editor of the new joumalActa Mathematica. She was the liaison with the mathematicians of Paris and Berlin. and took part in the organization of international conferences. Interestingly, Kovalevskaya also nurtured a parallel career in literature, penning several novels and a drama, "The Struggle for Happiness" that was favorably received at the Karsh Theater in Moscow. She died at the pinnacle

of her scientific careerfrom a combination of influenza and pneumonialess than twoyears after her election to both the Swedish and the Russian Academies of Sciences. The latter membership being initiated by Chebyshev, in spite of the Tsarist government's repeated refusal to grant her a university position in her own country.

21.1.2 Second-Order PDEs in m Dimensions Because of their importance in mathematical physics, the rest of this chapter and the next will be devoted to SOPDEs. This subsection classifies SOPDEs and the Bes associated with them .. The most general linear SOPDE in m variables can be written as

where A jk can be assumed to be symmetric in j and k. We restrict ourselves to the simpler case in which the matrix (A jk) is diagonal. We therefore consider the PDE (21.5) where the last term collects all the other terms except the second derivatives. We classify SOPDEs as follows: second-order POE of elliptic type second-order POEs of hyperbolic and ullrahyperbolic type second-order POE of parabolic type

1. Equation (21.5) is said to be of elliptic type at aj (xo) are nonzero and have the same sign.

xo if all the

coefficients

2. Equation (21.5) is said to be of ultrahyperbolic type at xo if all aj(xo) are nonzero but do not have the same sign. If only one of the coefficients has a sign different from the rest, the equation is said to be of hyperbolic type. 3. Equation (21.5) is said to be of parabolic type at coefficients aj (xo) is zero.

xo if at least one of the

If a SOPDE is of a given type at every point of its domain, it is said to be of that given type. In particular, if the coefficients aj are constants, the type of the PDE does not change from point to point.

590

21. MULTIDIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

21.1.8. Example. In this example, we study the SOPDE in two dimensions. The most general linear SOPDE is 2

2

au a2 u au ( au au) L[u]=a 2+2b--+c2+F x,y,u,;-,;- =0, ax axay ay ox oy

(21.6)

where a, b, and c are functions of x and y. To determine the characteristic curves of L, we seek conditions under which all secondorder partial derivatives of u can be determined from the DE and the Cauchy data, which are values of u and all its first derivatives on I', Consider a point Q == (x + dx, y + dy) close to P == (x, y). We can write

au

au

ax

ax

aZu ax z

axay

aZu

aZu axay

+ dy-(P), ayz

-(Q) - -(P) = d x - (PH dy-(P),

au

au

ay

ay

-(Q) - -(P) = dx-(P)

au

au)

a2u

-F (P, u(P), -(P), -(P) = a(P)2(P) ax ay ax

aZu

2u

2u

a a + 2b(P)""(P) +c(P)2(P), oxoy

ay

This system of three linear equations in the three unknowns-the three second derivatives evaluated at P-has a unique solution if and only if the determinant of the coefficients is nonzero. Thus, by Box 21.1.6, I' is a characteristic curve if and only if

det

dx 0 ( alP)

dy dx 2b(P)

~y) =0,

c(p)

or a(x, y)(dy)z - 2b(x, y)dxdy a(x, y) "" 0, that dy dx

ill-posed Cauchy problem

b±~ a

+ c(x, y)(dx)z

O. It then follows, assuming that

(21.7)

There are three cases to consider: 1. If b2 -ac < 0, Equation (21.7) hasno solution, which implies that no characteristic curves exist at P. Problem 21.1 shows that the SOPDE is of elliptic type. Thus, the Laplace equation in two dimensions is elliptic because b 2 - ac = -1. In fact, it is elliptic in the whole plane, or, stated differently, it has no characteristic curve in the entire xy-plane. This may lead us to believe that the Cauchy problem for the Laplace equation in two dimensions has a unique solution. However, even though the absence of a characteristic hypersurface at P is a necessary condition for the existence of a solution to the Cauchy problem, it is not sufficient. Problem 21.4 presents a Cauchy problem that is ill-posed, meaning that the solution at any fixed point is not a continuous function of the initial data. Satisfying this condition is required of a well-posed problem on both mathematical and physical grounds. 2. If bZ - ac > 0, Eqnation (21.7) has two solutions; that is, there are two characteristic

corvespassingthrongh P. Problem 21.1 showsthat the SOPDEis of hyperbolictype. The wave equation is such an equation in the entire ~2.

21.1 PROPERTIES OF PARTIAL OIFFERENTIAL EQUATIONS

591

3. If b 2 - ac = 0, Equation (21.7) has only one solution. In this case there is only one characteristic curve at P. The SOPDE is parabolic in this case. The one-dimensional

diffusion equation is an example of an SOPDE that is parabolic in the entire IRZ. IlIIII

appropriate BGs are determined bythe type of POE

Oirichlet boundary condition and boundary value problem

Neumann boundary condition and boundary value problem boundary conditions appropriate for elliptic, hyperbolic, and parabolic POEs

The question of what type of BCs to use to obtain a unique solution for a PDE is a very intricate mathematical problem. As Problem 21.4 shows, even though it has no characteristic curves in the entire R'", the two-dimensional Laplace eqnation does not lead to a well-posed Cauchy problem. On the other hand, examples in Chapter 19 that dealt with electrostatic potentials and temperatures led us to believe that a specification of the solution u on a closed curve in 2D, and a closed surface in 3D, gives a unique solution. This has a sound physical basis. After all, specifying the temperature (or electrostatic potential) on a closed surface should be enough to give us information about the temperature (or electrostatic potential) in the region close to the curve. A boundary condition in which the value of the solution is given on a closed hypersurface is called a Dirichlet boundary condition, and the associated problem, a Dirichlet HVP. There is another type of BC, which on physical grounds is appropriate for the Laplace equation. This condition is based on the fact that if the surface charge on a conductor is specified, then the electrostatic potential in the vicinity of the conductor can be determined uniquely. The surface charge on a conductor is proportional to the value of the electric field on the conductor. The electric field, on the other hand, is the normal derivative of the potential. A boundary condition in which the value of the normal derivative of the solution is specified on a closed curve is called a Neumann BC, and the associated problem, a Neumann bonndary value problem. Thus, at least on physical grounds, either a Dirichlet BVP or a Neumann BVP is a well-posed problem for the Laplace equation. For the heat (or diffusion) equation we are given an initial temperature distribution f(x) on a bar along, say the x-axis, with end points held at constant temperatures. For a bar with end points at x = a and x = b, this is equivalent to the data u(O, x) = f(x), uit; a) = Tj, and u(t, b) = h These are not Cauchy data, so we need not worry about characteristic curves. The boundary curve consists of three parts: (I) t = 0 for a :s: x :s: b, (2) t > 0 for x = a, and (3) t > 0, for x = b. In the xt-plane, these form an open rectangle consisting of ab as one side and vertical lines at a and b as the other two. The problem is to determine u on the side that closes the rectangle, that is, on the side a :s: x :s: b at t > O. The wave equation reqnires specification of both u and au/at at t = O. The displacement of the boundaries of the waving medium-a taut rope for examplemust also be specified. Again the curve is open, as for the diffusion case, but the initial data are Cauchy. Thus, for the wave equation we do have a Cauchy problem with Cauchy data specified on an open curve. Since the curve, the open rectangle, is not a characteristic curve of the wave equation, the Cauchy problem is wellposed. We can generalize these BCs to m dimensions and make the following correspondence between a SOPDE with m variables and the appropriate BCs:

I. Elliptic SOPDE +> Dirichlet or Neumann BCs on a closed hypersurface.

592

21. MULTIDIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

2. Hyperbolic SOPDE <+ Cauchy data on an open hypersurface.

3. Parabolic SOPDE <+ Dirichlet or Neumann BCs on an open hypersurface.

21.2 Multidimensional GFs and Delta Functions This section will discuss some of the characteristics of Green's functions in higher dimensions. These characteristics are related to the formal partial differential operator associated with the Green's function and also to the delta functions. Using the formal idea of several continuous indices, we can tum the operator equation LG = 1 into the PDE 8(x - y) LxG(x, y) =

w(x)

(21.8)

,

where x, y E jRm, w(x) is a weight function that is usually set equal to one, and,

only in Cartesian coordinates, m

8(x - y)

= 8(xl -

YI)8(xz - yz) ... 8(x m - Ym)

= T18(Xi i=l

- Yd·

(21.9)

In most applications Cartesian coordinates are not the most convenient to use. Therefore, it is helpful to express Equations (21.8) and (21.9) in other coordinate syslems. In particular, it is helpful to know how the delta function transforms under a general coordinate transformation. Let Xi = fi(l;l, ... , I;m), i = 1,2, ... , m, he a coordinate transformation. Let P he a point whose coordinates are a = (al, ... , am) and a = (al, ... , am) in the x and I; coordinate systems, respectively. Let J be the Jacobian of the transformation, that is, the absolute value of the determinant of a matrix whose elements are axilal;i- For a function F (x) the definition of the delta function gives Jd mxF(x)8(x-a) = F(a).Expressingthisequationintermsofthel; coordinate system, recalling that dmx = J dml; and a: = fiCa), and introducing the notation H(e) == F(fl(e), ... , fm(e)), we obtain

f singular point ofa !ranstorma!ion

dmUH(e)

fI

8(!i(e) - !i(a)) = H(a).

(21.10)

i=l

This suggests that J f17'~18(!i(e)- !i(a)) = f17'~18(l;i-ai),or,inmorecompact notation, J8(x - a) = 8(e - a). It is, of course, understood that J f= 0 at P. What happens when J = 0 at P? A point at which the Jacobian vanishes is called a singular point of the transformation. Thus, all points on the z-axis, including the origin, are singular points of Cartesian-spherical transformation. Since J is a determinant, its vanishing at a point signals lack of invertibility at that point.

593

21.2 MULTIDIMENSIDNAL GFS AND DELTA FUNCTIONS

ignorable coordinates

Thus, in the transformation from Cartesian to spherical coordinates, all spherical coordinates (5, tt, rp), with arbitrary tp, are mapped to the Cartesian coordinates (0,0, -5). Similarly, the point (0,0,0) in the Cartesian coordinate system goes to (0, e, rp) in the spherical system, with e and rp arbitrary. A coordinate whose value is not determined at a singular point is called an ignorable coordinate at that point. Thus, at the origin both e and rp are ignorable. Among the ~ coordinates, let {~i li'=k+l be ignorable at P with Cartesian coordinates a. This means that any function, when expressed in terms of ~ 's, will be independent of the ignorable coordinates. A reexamination of Equation (21.10) reveals that (see Problem 21.8)

I 8(x - a) =

-IJI

n8(~i k

- ail,

where

1k =

f Jd~k+l'" d~m'

(21.11)

k 1=1

In particular, ifthe transformation is invertible, k = m and Jm = J, and we recover J8(x - a) = 8(e - a).

21.2.1. Example. In two dimensions the transformation between Cartesian and polar == x = r eosB es gl COs;z. x2 == Y =7 sinO == ;1 singz with theJacobian

coordinates is given by Xl

which vanishes at the origin. The angle f:} is the only ignorable coordinate at the origin. Thus,k=2-1 = I, and

!) =

10r: J de = 10r: r de = 21fr =>

8(x)

== 8(x)8(y)

= -8(r) 2. 1fr

In three dimensions, the transformation between Cartesian and spherical coordinates yields Jacobian J = r 2 sin().Thisvanishes at the origin regardless of the values of 8 and tp, We thus havetwo ignorable coordinates at the origin (therefore, k = 3 - 2 = 1), over which we integrate to obtain

=> 21.2.1

8(r) 8(x) = 41fr 2'

III

Spherical Coordinates in m Dimensions

In discussing Green's functions in m dimensions, a particular curvilinear coordinate system will prove useful. This system is the generalization of spherical coordinates in three dimensions. The m-dimensional spherical coordinate system

594

21. MULTIDIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

is defined as Xl =

r sine!

Sillem-I,

X2 =

r sin (11

sinem _ 2 cos em-I, (21.12)

Xk =

r sin l:h

... sin {}m-k cos (}m-k+I,

2::s: k::s: m - I,

X m =rcos(h.

(Note that for m = 3, the first two Cartesian coordinates are switched compared to their nsual definitions.) It is not hard to show (see Example 21.2.2) that the Jacobian of the transformation (21.12) is a_)m-3 sm 01 )m-2(. SID V:l J :::::: r m-t(.

. ••

(si sm 0k )m-k-l

...

. 0m~2 sm

(21.13)

andthatthe volume element in termsof these coordinates is dmx = J dr dOl'" dOm-l = rm-ldr dQ.m, element ofthe m-dimensional solid angle

(21.14)

where dQ.m = (sinOl)m-2(sin!Jz)m-3 ... sinOm_2dOld02'" dOm-l

(21.15)

is the element of the m-dimensional solid angle. 21.2.2. Example. For m Xl = r

x3

= 4 we have

sin ~h sin Bz sin 93,

= r sin (h cos Oz.

X2

x4

= r sin 61 sin 8zcos 63, = rcos~h,

andtheJacobian is givenby

It is readily seen (one Can use mathematical induction to prove it rigorously) that the Jacobiaos for m = 2 (J = r), m = 3 (J = r 2 Sinel), aod m = 4 (J = r 3 sin2 el sine2) generalize to Equation (21.13). III

Jf:

Using the integral sin" 0 dO = .Jii'r[(n solid angle in m dimensions can be found to be

2n m / 2 Q. -:::-:---:;::m - r(m/2)'

+ 1)/2]/ r[(n + 2)/2], the total (21.16)

21.2 MULTIDIMENSIONAL GFS AND DELTA FUNCTIONS

595

An interesting result that is readily obtained is an expression of the delta function in terms of spherical coordinates at the origin. Since r = 0, Equation (21.12) shows that all the angles are ignorable. Thus, we have

which yields (21.17)

21.2.2 Green's Function for the Laplacian With the machinery developed above, we can easily obtain the (indefinite) Green's function for the Laplacian in m dimensions. We will ignore questions of Bes and simply develop a function that satisfies '1 2 G (x, y) = 8(x - y). Without loss of generality we let y = 0; that is, we translate the axes so that y becomes the new origin. Then we have '1 2 G (x) = 8(x). In spherical coordinates this becomes

8(r)

2

V G(x) =

Qm r m-

(21.18)

I'

Since the RHS is a function of r only, we expect G to behave in the same way. We now have to express '1 2 in terms of spherical coordinates. In general, this is difficult; however, for a function of r = have aF aXi

aF ar

aF Xi

= a;: aXi = a;:-;:

J

xf

+ ... + x~ alone, such as F(r), we

and

so that

For the Green's function, therefore, we get

~ (r m- I dr

dG) = 8(r). dr

(21.19)

Qm

The solution, for m :::: 3, is (see Problem 21.9) G(r) =

f(m/2) 2(m - 2)rr m / 2

(I) r m- 2

for

m:::: 3.

(21.20)

596

21. MULTIDIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

Green's function for the Laplacian

We can restore the vector y, at which we placed the origin, by notiog that r = [r] = [x - YI· Thus, we get r(m/2) ( I ) /2 2 2(m - 2)Jt m [x _ ylm-

G(x, y) =

r(m/2)

[

= - 2(m - 2)Jtm/2

2]-
t; m

for m ::: 3.

(Xi - Yi)

(21.21)

Similarly, for m = 2 we obtain G(x, y) solution of Poisson equation in m dimensions

I = -10 [x 217:

YI

I = -1o[(Xl 417:

Yl)

2

2

+ (X2 -

Y2) ].

(21.22)

Having found the Green's function for the Laplacian, we can find a solution to the inhomogeneous equation, the Poisson equation, V 2u = -pix). Thus, for m 2: 3, we get u(x)

=-

f f

m

(

d yG(x, y)p y)

r(m/2) _ 2)Jtm/2

= 2(m

f

m

plY)

d Y [x _ ylm-2'

In par1icular, for m = 3, we obtaio u(x) = _I

417:

d 3 y plY) ,

Ix- YI

which is the electrostatic potential due to a charge density ply).

21.3 Formal Development The precediog section was devoted to a discussion of the Green's function for the Laplacian with no mention of the BCs. This section will develop a formalism that not ouly works for more general operators, but also iocorporates the BCs.

21.3.1 General Properties Basic to a study of GFs is Green's identity, whose l-dimensional version we encountered io Chapter 20. Here, we generalize it to m dimensions. Suppose there exist two differential operators, Lx and Lk, which for any two functions u and v, satisfy the following relation: v*Lx[uJ - U(L~[V])* = V . Q[u, v*J sa

aQ' L -' [u, v*J. m

(21.23)

i=l aXi

The differential operator L~ is-as io the one-dimensional case--called the formal adjoint of Lx. Integrating (21.23) over a closed domaio D io R m with boundary

21.3 FORMAL DEVELOPMENT

597

aD, and using the divergeuce theorem, we obtain

1 D

generalized Green's identity

d"'x{v*Lx[u] -

u(~[v])*} =

(

laD

Q.

ell da,

(21.24)

where en is an m-dimensional unit vector normal to aD, and da is an etement of "area" of the m-dimensional hypersnrface aD. Equation (21.24) is the generalized Green's identity for m dimensions. Note that the weight function is set equal to one for simplicity. The differential operator Lx is said to be formally self-adjoint if the RHS of Equation (21.24), the snrface term, vanishes. In such a case, we have Lx = as in one dimension. This relation is a necessary condition for the surface term to vanish because u and v are, by assumption, arbitrary. Lx is called self-adjoint (or, somewhat imprecisely, hermitian) if Lx = L~ and the domains of the two operators, as determined by the vanishing of the snrface term, are identical. We can use Equation (21.24) to study the pair of PDEs

L!

Lx[u] = f(x)

L~[v] = h(x).

and

(21.25)

As in one dimeusion, we let G(x, y) and g(x, y) denote the Green's functions for Lx and Lt respectively. Let us assume that the Bes are such that the snrface term in Equation (21.24) vanishes. Then we get Green's identity

Iv

d"'xv*Lx[u] =

Iv d"'xu(~[v])*.

If in this equation we let u

Iv Green's functions are

symmetric functions oftheir arguments

d"'xg*(x,

y)~(x -

(21.26)

= G(x, t) and v = g(x, y), where t, y ED, we obtain

t) =

Iv

dmxG(x,

t)~(x -

y),

or g*(t, y) = G(y, t). In particular, when Lx is formally self-adjoint, we have G*(t, y) = G(y, t), or G(t, y) = G(y, t), if all the coefficient functions of Lx are real. That is, the Green's function will be symmetric. If we tet v = g(x, y) and use the first equation of (21.25) in (21.26), we get u(y) = fD d"'xg*(x, y)f(x), which, using g*(t, y) = G(y, t) and interchanging x and y, becomes u(x) = fD d"'yG(x, y)f(y). It can similarty be shown that v(x) = d"'yg(x, y)h(y).



21.3.2 Fundamental (Singular) Solutions The inhomogeneous term of the differential equation to which G (x, y) is a solution is the delta function, ~ (x - y). It would be surprising if G (x, y) did not "take notice" of this catastrophic source term and did not adapt itself to behave differently at x = y than at any other "ordinary" point. We noted the singular behavior of the Green's function at x = y in one dimension when we proved Theorem 20.3.4.

590

21. MULTIDIMENSIDNAL GREEN'S FUNCTIDNS: FDRMALISM

There we introduced h(x, y)-which was discoutinuous at x = y-as a part of the Green's function. Similarly, when we discussed the Green's functions for the Laplacian in two and m dimensions earlier in this chapter, we noted that they behaved singularly at r = 0 or x = y. In this section, we study similar properties of the GFs for other differential operators. Next to the Laplacian in difficulty is the formally self-adjoint elliptic PD~ Lx = V2 + q(x) discussed in Problem 21.10. Substituting this operator in the generalized Green's identity and using the expression for Q given in Problem 21.10, we obtain [ dmx{vLx[u] - u(Lx[v])} = [

Jv

laD

Letting v = G(x, y) and denoting

en' V by a/an gives

[ dmx[GL,.u _ uLxG] = [

JD

(ve n • Vii - ue n . Vv) da.

Jan

[G

au an

_ u aG] da. an

(21.27)

We wantto use this equationto find out about the behaviorof G(x, y)as Ix-YI ~ O. Therefore, assuming that y ED, we divide the domain D into two parts: one part is a region DE bounded by an infinitesimal hypersphere SE with radius E and center at y; the other is the rest of D. Instead of D we use the region D' sa D - DE' The following facts are easily deduced for D': (1) LxG(x, y) = 0 because x yin D'; (2) = limHo (3) aD' = eo USE' Suppose that we are interested in finding a solution to

JD

t=

JD';

Lx[U] = [V 2 + q(x)]u(x)

= f(x)

subject to certain, as yet unspecified, BCs. Using the three facts listed above, Equation (21.27) yields [ dmx[GL,.u - uLxG] = lim [

J

J.D

dmx[G Lxu -u LxG]

E--+O D'

= lim [

-----"-.,-'

=1

=0

dmxG(x, y)f(x) = [ dmxG(x, y)f(x)

Jv

E--+O}DI

= [

Jan

(G

au

_u

aG)

da+[ (G

au

_u

Js, an We assume that the BCs are such that the integral over aD vanishes. an

an

aG) da. an

This is a generalization of the one-dimensional case (recall from Chapter 20 that this is a necessary condition for the existence of Green's functions). Moreover, for an m-dimensional sphere, da = rm-1dfl m, which for SE reduces to Em-1dfl m. Substituting in the preceding equation yields [ dmxG(x, y)f(x) = [

JD

Js,

(G au _ u aG) Em-1dfl m. an an

21.3 FORMAL DEVELOPMENT

599

We would like the RHS to be u(y). This will be the case if and for arbitrary u. This will happen only if lim G(y + r, y)r..- 1 = 0,

r-+O

aG lim - ( y + r, y)r..- 1 = const.

r--+O

(21.28)

ar

A solution to these two equations is - F(x, y) Intjx _ yl) G(x, y) =

lundamental solution isthe singular part 01 GF

homogeneous solution is the regular part 01 GF

{

2" I

F(x, y)

C"(--2"")-=n-

m-

+ H(x, y)

;:"(,m

'1--'--',1"' x-y ..'-;.,2 +

if m = 2, (21.29)

H (x, y)

if m 2: 3,

where H(x, y) and F(x, y) are well behaved at x = y. The introduction of these functions is necessary because Equation (21.28) determines the behaviorof G(x, y) only whenx "" y. Such behavior does not uniquely determine G(x, y). For instance, elx-ylln(lx - YI) and Intlx - YI) behave in the same way as [x - YI ..... O. Equation (21.29) shows that for Lx = V 2 + q (x), the Green's function consists of two parts. The first part determines the singular behavior of the Green's function asx ..... y. The natore of this singularity (how badly the GF "blows up" asx ..... y)is extremely important, because it is a prerequisite for our ability to write the solution in terms of an integral representation with the Green's function as its kernel. Due to their importance in such representations, the first terms on the RHS of Equation (21.29) are called the fundamental solution of the differential equation, or the singular part of the Green's function. What about the second part of the Green's function? What role does it play in obtaining a solution? So far we have been avoiding consideration of BCs. Here H(x, y) can help. We choose H(x, y) in such a way that G(x, y) satisfies the appropriate BCs. Let us discuss this in greater detail and generality. If BCs are iguored, the Green's function for a SOPDO Lxcannot be determined uniquely. In particular, if G(x, y) is a Green's function, that is, if LxG(x, y) = 8(x - y), then so is G(x, y) + H(x, y) as long as H(x, y) is a solution of the homogeneous equation LxH(x, y) = O. Thus, we can break the Green's function into two parts:

G=G,+H,

where

LxG,(x, y)

= 8(x -

y),

LxH(x, y)

=0 (21.30)

regular part 01 the Green's function

with G, the singular part of the Green's function. H is called the regular part of the Green's function. Neither G, nor H (nor G, therefore) is unique. However, the appropriate BCs, which depend on the type of Lx, will determine G uniquely. To be more specific, let us assume that we want to find a Green's function for Lx that vanishes at the boundary aD. That is, we wish to find G(x, y) such that

600

21. MULTIDIMENSIDNAL GREEN'S FUNCTIDNS: FDRMALISM

G(Xb, y) = 0, where Xb is an arbitrary point of the boundary. All that is required is to find a G, and an H satisfying Equatiou (21.30) with the BC H(Xb, y) = -G,(Xb, y). The latter problem, involving a homogeneous differential equation, can be handled by the methods of Chapters 18 and 19. Since any discussion of BCs is tied to the type of PDE, we have reserved the discussion of such specifics for the next chapter.

21.4 Integral Equations and GFs Integral equations are best applied in combination with Green's functions. In fact, we can use a Green's function to tum a DE into an integral equation. If this integral equation is compact or has a compact resolvent, then the problem lends itself to the methods described in Chapters 16 and 17. Let Lx be a SOPDO in m variables. We are interested in solving the SOPDE Lx[u] + J..V(x)u(x) = fix) snbject to some BCs. Here J.. is an arbitrary constant, and V (x) is a well-behaved function on jRm. Transferring the second term on the LHS to the RHS and then treating the RHS as an inhomogeneous term, we can

write the "solution" to the PDE as u(X) = H(x)

+

L

dmyGo(x, y)[f(y) - AV(y)u(y)],

where D is the domain of Lx and Go is the Green's functiou for Lx with some, as yet unspecified, BCs. The functiou H is a solution to the homogeneous equation, and it is present to guarantee the appropriate BCs. Combining the first term iu the integral with H(x), we have

L

u(x) = F(x) - J..

dmyGo(x, y)V(y)u(y).

(21.31)

Equation (21.31) is an m-dimeusioual Fredholm equation whose solution can be obtained in the form of a Neumann series. 21.4.1. Example. Consider thebound-state Schriidinger equation in one dimension: 1i2 d 2'11

--+ V(x)'II(x) 2 2", dx

= E'II(x),

E
Werewrite thisequation as d

2

Lx['II] '" ( dx 2 -

where K 2

K

2)

2",

'II(x) = 1i2 V(x)'II(x),

= -2",E/1i2 > O. Equation (21.31) givesthe equivalent integralequation

'II(x) = 'IIo(x)

2", + 2"

Ii

joo -00

Oo(x, y)V(y)'II(y)dy

601

21.4 INTEGRAL EQUATIONS AND GFS

= 0, whichis easily found to be of the general form "Wo(x) = Ae lCx + Be-KX • If we assume that \I1o(x) remains finite as x -....+ ±oo, wo(x) will be zero. Furthermore, it can be shown that Go(x, y) = -e-Klx-yI/2K (see Problem

where Wo(x) is the solution of Lx [Wol

20.12). Therefore, W(x) =

-4foo fi 1C

e-Klx-YIY(y)W(y)dy.

-00

Now consider anattractive delta-function potential with center at a: V (x)

= -

Vo8 (x -

a), Vo > O. Forsucha potential, the integral equation yields

= ..!!:.- fOO

W(x)

~K

e-Klx-YIYo8(y _ a)W(y) dy

= '"' YO e-Klx-aIW(a).

-00

~K

Forthis equationto be consistent, i.e., to get an identitywhen x = a, we must have

,",Yo

lih

,",Yo =1,* K=1iJ

,",Yo '* E=-21i2'

Therefore, there is onlyoneboundstateandoneenergylevel foranattractive delta-function

potential.

III

Tofinda Neumann-series solutionwe cansubstitute theexpression for u given by the RHS of Equation (21.31) in the integral of that equation. The resulting equation will have two integrals, in the second of which u appears. Substituting the new u in the second integral and continuing the process N times yields N-l

(

n=l

D

I)-A)n J,

u(x) = F(x)+

+ (_A)N

Iv

dmyKn(x, y)F(y)

dmyK N (x, y)u(y),

where K(x, y) es Y(x)Go(x, y), Kn(x, y)

==

Iv

dmtKn-t(x, t)K(t, y)

for n 2: 2. (21.32)

The Neumano series is obtained by lettiog N -> u(x) = F(x)

+ f(-A)n n=l

1

00:

dmyKn(x, y)F(y).

(21.33)

D

Except for the fact that here the integrations are in m variables, Equation (21.33) is the same as the Neumano series derived in Section 17.1. In exact analogy, therefore, we abbreviate (21.33) as 00

lu) = IF)

+ L(-A)nKn IF). n=l

(21.34)

602

21. MULTIDIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

Equations (21.33) and (21.34) have meaning only if the Neumann series converges, i.e., if (21.35)

Feynman's diagrammatic representation of GF

We will briefly discuss an intuitive physical interpretation of the Neumann series due to Feymuan. Although Feymuan developed this diagrammatic technique for quantum electrodynamics, it has been useful in other areas, such as statistical and condensed matter physics. In most cases of interest, the SOPDE is homogeneous, so f(x) = O. In that case, .... and V(x) are called the free operator and the interacting potential, respectively. The solution to Lx[u] = 0 is called the free solution and denoted by u f(x). Let us start with Equation (21.31) written as u(x) = uf(x) - A [

dmyGo(x, y)V(y)u(y),

JJRIll

(21.36)

where Go stands for the Green's function for the free operator Lx. The full Green's function, that is, that for Lx+ AV, will be denoted by G. Moreover, as is usually the case, the region D has been taken to be all of Em. This implies that no boundary conditions are imposed on u, which in turn permits us to use the singular part of the Green's function in the integral. Because of the importance of the full Green's function, we are interested in finding a series for G in terms of Go, which is supposed to be known. To obtain such a series we start with the abstract operator equation and write G = Go + A, where A is to be determined. Operating on both sides with L ("inverse" of Go), we obtain LG = LGo + LA = 1 + LA. On the other hand, (L + AV)G = 1, or LG = 1 - AVG. These two equations give LA = -AVG =} A = -AL-IVG = -AGoVG.

Therefore, G = Go - AGoVG.

(21.37)

Sandwiching both sides between (x] and [z), inserting 1 = fly) (YI dmy between Go and V and 1 = fit) (tl dmt between V and G, and assuming that V is local [i.e., V(y, t) = V(y)8(y - t)], we obtain G(x,z) = Go(x,z) -A

f

dmyGo(x,y)V(Y)G(y,z).

(21.38)

This equation is the analogue of (21.31) and, just like that equation, is amenable

to a Neumann seriesexpansion. Theresultis G(x, y) = Go(x, y)

+ I)-A)n 1m n=l

JRIll

dmzGo(x, z)Kn(z, y),

(21.39)

21.5 PERTURBATION THEORY

~

603

y

Yz

YI

X

(a)

(b)

(c)

Figure21.1 Contributions to the full propagatorin (a) the zerothorder, (b) the firstorder, and (c) thesecond order. At eachvertex oneintroduces a factor of -}.. V andintegrates over all values of the variable of that vertex.

where K" (x, z) is as given in Equation (ZI.32). Feynman's idea is to consider G(x, y) as an interacting propagator between points x and y and Go(x, y) as a free propagator. The first term on the RHS of (21.39) is simply a free propagation from x to y. Diagrammatically, it is represented by a line joining the points x and y [see Figure 21.1(a)]. The second term is a free propagation from x to YI (also called a vertex), interaction at YI with a potential -A V (y I), and subsequent free propagation to y [see Figure 21.1 (b)]. According to the third term, the particle or wave [represented by U f (x)] propagates freely from x to YI, interacts at YI with the potential-AV(YI), propagates freely from YI to Y2, interacts for a second time with the potential -A V(Y2), and finally propagates freely from Y2 to y [Figure 21.1(c)]. The interpretation of the rest of the series in (21.39) is now clear: The nth-order term of the series has n vertices between x and Y with a factor -AV(Yk) and an integration over Yk at vertex k. Between any two consecutive vertices Yk and Yk+I there is a factor of the free propagator GO(Yk, Yk+l). Feynman diagrams are used extensively in relativistic quantum field theory, for which m = 4, corresponding to the four-dimensional space-time. In this context A is determined by the strength of the interaction. For quantum electrodynamics, for instance, A is the fine-structure constant, e 2 /1i.c = 1/137.

21.5 Perturbation Theory Few operator equations lend themselves to an exact solution, and due to the urgency of findiug a solution to such equations in fundamental physics, various techniques have been developed to approximate solutions to operator equations. We have already seen instances of such techniques in, for example, the WKB method. This section is devoted to a systematic development of perturbation theory, which is one of the main tools of calculation in quantum mechartics. For a thorough treatment of perturbation theory along the lines presented here, see [Mess 66, pp. 712-720].

604

21. MULTIDIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

The starting point is the resolvent (Definition 16.8.1) of a Hamiltonian H, which, using z instead of A, we write as Rz(H). For simplicity, we assume that the eigenvalues of H are discrete. This is a valid assumption if the Hamiltonian is compact or if we are interested in approximations close to one of the discrete eigenvalues. Denoting the eigenvalues of H by {Ed;::;o, we have (21.40) where Pi is the projection operator to the ith eigenspace. We can write the resolvent in terms of the projection operators by using Equation (16.8): Pi

L--' i=O Ei - Z 00

Rz(H) =

(21.41)

The projection operator Pi can be written as a contour integral as in Equation (16.13). Any sum of these operators can also be written as a contour integral. For instance, if r is a circle enclosing the first n + 1 eigenvalues, then Pr

n == LPi

= - -1.

i

2.7rl

J'

i=O

Rz(H)dz.

(21.42)

Multiplying Equation (21.42) by H and using the definition of the resolvent, one can show that 1

HPr = - 2 .

m

perturbing potential

J zRz(H) dz: ir

(21.43)

When r includes all eigenvalues of H, Pr = 1, and Equation (21.43) reduces to (16.12) with A --+ T and f(x) --+ x. To proceed, let us assume that H = Ho + AV where Ho is a Hamiltonian with known eigenvalues and eigenvectors, and V is aperturbing potential; Ais a (small) parameter that keeps track of the order. of approximation. Let us also use the abbreviations G(z) es -Rz(H)

and

Go(z) sa -Rz(Ho).

(21.44)

Then a procedure very similar to that leading to Equation (21.37) yields G(z) = Go(z)

+ AGo(z)VG(z),

(21.45)

which can be expanded in a Neumano series by iteration: 00

G(z) = LAnGO(zHVGo(z)]". n=oO

(21.46)

Let {E~l, {M~ I, and rna denote, respectively, the eigenvalnes of Ho, their corresponding eigenspaces, and the latter's dimensions.e In the context of pertorbation

605

21.5 PERTURBATION THEORY

Degeneracy isthe dimension ofthe eigenspace ofthe Hamiltonian.

E2,

E2

theory, rna is called the degeneracy of and is called rna-fold degenerate, with a similar terminology for the perturbed Hamiltonian. We assnme that all

eigenspaces havefinite dimensions. It is clear that eigenvalnes and eigenspaces of H will tend to those of Ho when

M2

A --> O. So, let us collect all eigenspaces of H that tend to and denote them by {Jv(f};~1' Similarly, we use Ei andPf to denote, respectively, the energy eigenvalue and the projector to the eigenspace Mi. Since dimension is a discrete quantity, it cannot depend on A, and we have

L'" dim Mi

=

dimM~ = rna.

(21.47)

i=l

We also use the notation P for the projector onto the direct snm of Mi's. We thus have and

HmP=P~,

(21.48)

).-->0

M2.

where we have used an obvious notation for the projection operator onto The main task of perturbation theory is to find the eigenvalues and eigenvectors of the perturbed Hamiltonian io terms of a series io powers of A of the corresponding unperturbed quantities. Sioce the eigenvectors-c-or, more appropriately, the projectors onto eigenspaces-and their corresponding eigenvalues of the perturbed Hamiltonian are related via Equation (21.40), this task reduces to writing P as a series in powers of A whose coefficients are operators expressible io terms of unperturbed quantities. For sufficiently small A, there exists a contour io the z-plane enclosiog and all E't's but excludiog all other eigenvalues of H and Ho. Denote this contour by r a and, nsiog Equation (21.42), write

E2

P=

~J 2nl

ira G(z)dz.

It follows from Equation (21.46) that 00

p=p2+ I>nA(n), Il=l

where

A(n)

==

_1_. 21r1

J

Yr"

Go(z)[VGo(z)]" dz: (21.49)

This equation shows that perturbation expansion is reduced to the calculation of A(n), which is simply the residue of Go(z)[VGo(z)]n. The only siogularity of the iotegrand in Equation (21.49) comes from Go(z), which, by (21.44) and (21.41), 2Weuse thebeginningletters of theLatinalphabet for theunperturbed Hamiltonian. Furthermore, we attach a superscript "0" to emphasizethatthe objectbelongs to HO'

606

21. MULTIDIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

has a pole at E2. So, to calculate this residue, we simply expaod Go(z) in a Laurent series about

Eg: pO

~=

Go(z) = L

b Z - Eb

=

°

Pa ---0 z - Ea

pO

_ _a_ Z - EO a

-

pO _a_ o Z - Ea

pO

~

+L

b'!a Z - E b

pO b

~

+ L."

b'!a (Eo _ EO) a b

+ b¥=ak=O L." L."

°

z - Ea ) E2- E

g

(z _ EO)kpO

00

~~(_I)k

(I +

a

b

(EO _ EO)k+1 . a

b

Switching the order of the two sums, aod noting that our space is the Hilbert space of Ho whose basis cao be chosen to consist of eigenstates of Ho, we cao write Ho instead of in the denominator to obtain

Eg

where we have used the completeness relation for the Pg's, the fact that Q2 commutes with Ho [aod, therefore, with G~+l (E2)] , aod, in the last equality, the fact that Q2 is a projection operator.' It follows that Go(z) =

7pO Z -

00

+ L(-Ihz a

E2)kQ2G~+1(E~)Q~

k=O

00

= L(-lhz - E~)k-lSk, k=O

(21.50)

where we have introduced the notation

if k = 0, if

k:o>:1.

3Note that although Go (z) has a pole at E~. the expressions in the last line of the equation abovemakesense because Q~ annihilates all stateswith eigenvalue E~. Thereasonfor theintroduction of Q~ on bothsides is to ensurethat G~+ 1(E~) will not acton aneigenstate of E~ on either side.

607

21.5 PERTURBATION THEORY

By substituting Equation (21.50) in Go(z)[VGo(z)]n we obtain a Laurent expansion whose coefficient of (z - E2)-1 is A(n). The reader may check that such a procedure yields A(n)

= (_1)"+1 LSk1VSk'V ... VSk"+l,

(21.51)

(n)

where by definition,

L(p)

extends over all nonnegative integers {ki }?~l such that

n+l

Lki=p

v p ::::0.

i=l

It turns out that for perturbatiou expansion, uot ouly do we need the expansion of P [Equations (21.49) and (21.51)], but also an expansion for HP. Using Equations (21.43) and (21.44), with r replaced by r a, we have HP

=~J

J (z fta zG(z)dz = ---.!:..., 27fl jr a ---.!:..., J (z - E2)G(z) dz + E2p. 2Jrl fta

E2

21ft

=

+ E2)G(z)dz

Substituting for G(z) from Equation (21.46), we can rewrite this equation as 00

(H - E2)p = L}.nB(n),

(21.52)

n=l

where B(n)

= (_l)n-l L

Sk' VSk'V· .. VS k"+l

(21.53)

(n-l)

Equations (21.52) and (21.53) can be used to approximate the eigenvectors and eigenvalues of the perturbed Hantiltonian in terms of those of the unperturbed Hamiltonian. It is convenient to consider two cases: the nondegenerate case in which rna = 1, and the degenerate case in which rna :::: 2.

21.5.1 The Nondegenerate Case In the nondegenerate case, we let 12) denote the original unperturbed eigenstate, and use Equation (21.47) to conclude that the perturbed eigenstate is also onedimensional. In fact, it follows from (21.40) that P I~) is the desired eigenstate. Denoting the latter by Ilfr) and using Equation (21.49), we have 00

Ilfr)

= P 12) =

p212) + L}.nA(n) I~) n=l

00

= I~)

+ L}.nA(n) 12) n=l

(21.54)

608

21. MULTIDIMENSIDNAL GREEN'S FUNCTIDNS: FDRMALISM

because P~ is the projection operator onto I~). More desirable is the energy of the perturbed state E a , which obeys the relation HP = EaP. Taking the trace of this relation and noting that tr P = tr P~ = I, we obtain

e, = tr(HP) = tr (E~P + ~ AnB(n») (21.55) where we used Equation (21.52). Since A is sinnply a parameter to keep track of the order of perturbation, one usually includes it in the definition of the perturbing potential v. The nth-order correction to the energy is then written as

en =

(21.56)

tr B(n) .

Since each term of B(n) contains p~ at least once, and since

tr(UP~T) = tr(TUP~)

first-order correction to energy

for any pair of operators U and T (or products thereof). one can cast en into the form of an expectation value of some product of operators in the unperturbed state I~). For example, el

= tr B(I) = L: (g I P~VP~ Ig) = (~I V I~)

(21.57)

b

=

=

because P~ Ig) 0 unless b a. This is the familiar expression for the first order correction to the energy in nondegenerale perturbation theory. Similarly,

ez =

tr B(2) --

-

tr (poVpoV[-OoGk(Eo)Oo] aa aQaa

o + [_OoGk(Eo)Oo]VpoVpo) +poV[_OoGk(Eo)Oo]Vp a aOaaa aOaaaa = (~I VO~G~(E~)O~V I~).

The first and the last terms in parentheses give zero because in the trace sum, p~ gives a nonzero contribution only if the state is I~), which is precisely the stale anniltilated by O~. Using the completeness relation Lb Ig) (gl = 1 = L c I~) (~I

second-order correction to energy

for the eigenstates of the unperturbed Hamiltonian, we can rewrile ez as

This is the familiar expression for the second-order correction to the energy in nondegenerate perturbation theory.

21.5 PERTURBATION THEORY

609

21.5.2 The Degenerate Case The degenerate case can also start with Equations (21.54) and (21.55). The difference is that en cannot be determined as conveniently as the nondegenerate case. Forexample,the expression for 81 will involve a sum over a basis of Jv(~ because P~ Ig} is no longer just I~)' but some general vector in JY[~. Instead of pursuing this line of approach, we presenta more common method, which concentrates on the way JY[~ and the corresponding eigenspaces of the perturbed Hamiltonian, denoted by JY[a, enter in the calculation of eigenvalues and eigenvectors. The projector P~ acts as a unit operator when restricted to JY[~. In particular, it is invertible. In the limit of small A,the projection operator P is close to P~; therefore, it too must be invertible, i.e., P : JY[~ -+ JY[a is an isomorphism. Similarly, P~ : JY[a -+ JY[~ is also an isomorphism-not necessarily the inverse of the first one. It follows that for each vector in JY[~ there is a unique vector in JY[a and vice versa. The eigenvalue equation H lEa) = E a lEa) can thus be written as HPa IE2} = EaP. IE2) ,

where IE2) is the unique vector mapped onto IE.) by pa- Multiplying both sides by P~, we obtain P2HPa IE2} =

EaP~Pa IE2) ,

which is completely equivalent to the previous equation because P~ is invertible. If we define Ha

== P~HPaP~ : M~

-> JV(~,

(21.58)

the preceding equation becomes Ha IE2} = EaK" IE~}.

(21.59)

As operators on JY[~ both Ha and Ka are hermitian. In fact, K", which can be written as the product of P~Pa and its hermitian conjugate, is a positive definite operator. Equation (21.59) is a generalized eigenvalue equation whose eigenvalues Ea are solutions of the equation (21.60)

det (Ha - xK,,) = O.

The eigenvectors of this equation, once projected onto JY[a by P a, give the desired eigenvectors of H. The expansions of Ha and Ka are readily obtaiued from those of HPa and Pa as given in Equations (21.49) and (21.52). We give the first few terms of each expansion:

K" = ~ - A2~VQ~Gt(E~)Q~V~ + ... ,

°

°

°

o 'po.VPa+A,2 PaVaO QOG (EO)QO Ha = E aKa+A a aVPa+,.·· (21.61)

610

21. MULTIDIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

To any given order of approximation, the eigenvalues E a are obtained by terminating the series in (21.61) at that order, plugging the resulting finite sum in Equation (21.60), and solving the determinant equation.

21.6 Problems 21.1. Show that the definitions of the three types ofSOPDEs discussed in Example 21.1.8 are equivalent to the definitions based on Equation (21.5). Hint: Diagonalize the matrix of coefficients of the SOPDE: 2u

2u

a a a -2 + 2 baa ax x y

2u

a ( au au) +c+F x,y,u, -a '-a =0, 2 ay x y

where a, b, and c are functions ofx and y. Write the eigenvalues as (a +c± 1:;.)/2 and consider the three cases II:;.I < la + c], II:;.I > la + c], and II:;.I = la + c]. 21.2. Find the characteristic curves for Lx[u] = au/ax. 21.3. Find the characteristic curves for the two-dimensional wave equation and the two-dimensional diffusion equation. 21.4. Solve the Cauchy problem for the two-dimensional Laplace equation subject to the Cauchy data u(O, y) = 0, (au/ax) (0, y) = E sinky, where E and k are constants. Show that the solution does not vary continuously as the Cauchy data vary. In particular, show that for any E i' and any preassigned x > 0, the solution u(x, y) can he made arbitrarily large by choosing k large enongh.

°

21.5. Show that the

radius

Xi

in Equation (21.12) describe an m-dimensional sphere of

r, that is, Lr=l xl =

,2.

21.6. Use J8(x - a) = 8(e - a) and the coordinate transformation from the spherical coordinate system to Cartesian coordinates to express the 3D Cartesian delta function in terms of the corresponding spherical delta function at a point P = (xo, Yo, zo) = (ro, 1J0, 9'0) where the Jacobian J is nonvanishing, 21.7. Find the volmne of an m-dimensional sphere. 21.8. Prove Equation (21.11). First, note that the RHS of Equation (21.10) is a function of only k of the a's. This means that

(a) Rewrite Equation (21.10) by separating the integral into two parts, one involving {~;\~~1 and the other involving (~;\;n~k+l' Compare the RHS with the LHS and show that

JJd~k+l'" d~m8(x-a)

=

fI8(~i -ail·

i=l

21.6 PROBLEMS

611

(b) Show thatthis equation implies that 8(x - a) is independent of Rilf'~k+l' Thus, one can take the delta function out of the integral.

21.9. Find the m-dimensional Green's function for the Laplacian as follows. (a) Solve Equation (21.19) assuming that r oF 0 and demanding that G(r) -+ 0 as r -+ 00 (this can be done only for m 2: 3). (b) Use the divergence theorem in m dimensions and (21.18) to show that

II ~~

da= I,

s

where S is a spherical hypersurface of radius r. Now use this and the result of part (a) to find the remaining constant of integration.

21.10. Consider the operator Lx = V2 + b . V

+ c for which

{hilf'~l and c are

functions of {x; }f'~1' (a) Show that !J[v] = V 2 v - V . (bv)

Q[u, v*] = Q[u, v] = vVu -

+ cv, and uVv + buv.

(b) Show thatanecessary condition for Lxto be self-adjoint is Zb-Vu+u(V -b) = 0 for arbitrary u. (c) By choosing some u'sjudiciously, show that (b) implies that hi = O. Conclude that Lx = V 2 + c(x) is formally self-adjoint.

21.11. Solve the integral form of the Schrodinger equation for an attractive double delta-function potential

Vex) = -Vo[8(x - al) + 8(x - a2)],

Vo > O.

Find the eigenfunctions and obtain a transcendental equation for the eigenvalues (see Example 21.4.1).

21.12. Show that the integral equation associated with the damped harmonic oscillatorDEx +2yi + w5x = 0, having the BCs x(O) = xo, (dx/dt)t~o = 0, can be written in either of the following forms.

w21t (a) x(t) = Xo - --..!! 2y

°

[1- e- 2y (t- t')] x(t') dt'.

(h) x(t) = xocoswot + 2y

xO

wo

sinwot - 2y [' cos[wo(t - t')]x(t')dt'.

io

Hint: Take w5x or 2yi, respectively, as the inhomogeneous term.

21.13. (a) Show that for scattering problems (E > 0) the integral form of the Schrodinger equation in one dimension is ljI(x) = e;kx _

i~

fj k

1

00

-00

e;klx-yl V(y)ljI(y) dy.

612

21. MULTIDIMENSIONAL GREEN'S FUNCTIONS: FORMALISM

(b) Divide (-00, +00) into three regions Rl = (-00, -a), Rz = (-a, +a), and R3 = (a, 00). Let l!ri(X) be 'f!(x) in region Ri, Assume that the potential V(x) vanishes in R 1 and R3. Show that 'f!1(X) = e ' kx _

i': Ii k ~~

'f!z(x) = e ' kx -

e- ikx

i:

1 a

e'kYV(y)'f!z(y)dy,

-a

e'klx-yIV(y)'f!z(y)dy,

1 a

'f!3(X) = e ikx _ .!..!!:.-elkx e-lkYV(y)'f!z(y)dy. Itzk -a

This shows that determining the wave function inregions where there is no potential requires the wave function in the region where the potential acts. (c) Let V(x) =

{~o

if [x] < a,

if

Ixl

> a,

and find 'f!z(x) by the method of successive approximations. Show that the nth term is less than (21-' Voa/ItZk)n-l (so the Neumann series will converge) if (2 Voa/ ltv) < I, where v is the velocity and I-' v = Iik is the momentum of the wave. Therefore, for large velocities, the Neumann series expansion is valid. 21.14. (a) Show that HRz(H) = 1 +zRz(H). (b) Use (a) to prove Equation (21.43).

Additional Reading 1. Folland, G. Introduction to Partial Differential Equations, 2nded., Princeton University Press, 1995. Discusses multidimensional Green's functions for various differential operators of mathematical physics. 2. Messiah, A. Quantum Mechanics, volume II, Wiley, 1966. A thorough treatment of perturbation theory in the style of this chapter. 3. Stakgold.L Green's Functions and Boundary Value Problems, Wiley, 1979. A detailed analysis of boundary value problems in two and three dimensions.

22

_

Multidimensional Green's Functions: Applications

The previous chapter gathered together some general properties ofthe GFs and their companion, the Dirac delta function. This chapter considers the Green's functions for elliptic, parabolic, and hyperbolic equations that satisfy the BCs appropriate for each type of POE.

22.1 Elliptic Equations The most general linear POE in m variables of the elliptic type was discussed in Section 21.1.2. We will not discuss this general case, because all elliptic PDQs encountered in mathematical physics are of a much simpler nature. In fact, the self-adjoint elliptic PD~ of the form Lx = '1,72 + q(x) is sufficiently general for purposes of this discussion. Recall from Section 21.1.2 that the BCs associated with an elliptic POE are of two types, Dirichlet and Neumann. Let us consider these separately.

22.1.1

The Dirichlet Boundary Value Problem

A Dirichlet BVP consists of an elliptic POE together with a Dirichlet BC, such as Lx[u] = 'I,72 u

+ q(x)u =

U(Xb) = g(Xb)

f(x)

for Xb E aD,

for xED, (22.1)

where g(Xb) is a given function defined on the closed hypersurface aD. The Green's function for the Dirichlet BVP must satisfy the homogeneous BC, for the same reason as in the one-dimensional Green's function. Thus, the Dirichlet

614

22. MULTIDIMENSIONAL GREEN'S FUNCTIONS: APPLICATIONS

Green's function, denoted by Gn(x, y), must satisfy

for Xb E S.

Lx[Gn(x, y)] = 8(x - y),

As discussed in Section 21.3.2, we can separate Gn into a singular part Gg) and a regular part H where G

g) satisfies the same DE as G n and H satisfies the

corresponding homogeneous DE and the Be H(Xb, y) = -Gg) (Xb, y). Using Equation (22.1) and the properties of G n(x, y) in Equation (21.27), we obtain u(X) = ( d"'yGn(x, y)f(y)

JD

+(

laD

Gn g(Yb) aa (x, Yb) da,

ny

(22.2)

where ajan y indicates normal differentiation with respect to the second argument.

Gustav Peter Lejeune Dirichlet (1805-1859), thesonof apostmaster, firstattendedpublicschool,thenaprivateschoolthatemphasized Latin. He was precociously interested in mathematics; it is said that before the age of twelve he used his pocket money to buy mathematical books. In 1817 he entered the gymnasium in Bonn. He is reported to have been an unusually attentive and well-behaved pupil who was particularly interested in modem

historyas well as in mathematics. After two years in Bonn, Dirichlet was sent to a Jesuit college in Cologne that his parents preferred. Among his teachers was the physicist Georg Simon Ohm, who gave him a thorough grounding in theoretical physics. Dirichlet completed his Abitur examination at the very early age of sixteen. His parents wanted him to study law, but mathematics was already his chosen field. At the time the level of pure mathematics in the German universities was at a low ebb: Except for the formidable Carl Gauss, in G6ttingen, there were no outstanding mathematicians, while in Paris the firmament was studded by such luminaries asP.-S. Laplace, Adrien Legendre, Joseph Fourier, and Simeon Poisson. Dirichlet arrived in Paris in May 1822. In the summer of 1823 he was fortunate in being appointed to a well-paid and pleasant position as tutor to the children of General Maximilien Fay, a national hero of the Napoleonic wars and then the liberal leader of the opposition in the Chamber of Deputies. Dirichlet was treated as a member of the family and met many of the most prominent figures in French intellectual life. Among the mathematicians, he was particularly attracted to Fourier, whose ideas had a strong influence upon his later works on trigonometric series and mathematical physics. General Fay died in November 1825, and the next year Dirichlet decided to return to Germany, a plan strongly supported by Alexander von Humboldt, who was working for the strengthening of the natural sciences in Germany. Dirichlet was permitted to qualify for habilitation as Privatdozent at the University of Breslau; since he did not have the required doctorate, this was awarded honoris causa by the University of Cologne. His habilitation thesis dealt with polynomials whose prime divisors belong to special arithmetic series. A second paper from this period was inspired by Gauss's announcements on the biquadratic law of reciprocity.

22.1

ELLIPTIC EQUATIONS

615

Dirichlet was appointedextraordinaryprofessor in Breslau, but the conditionsfor scientific work were not inspiring. In 1828 he moved to Berlin, again with the assistance of

Humboldt.to becomea teacher of mathematicsat the military academy. Shortlyafterward, at the age of twenty-three, he was appointed extraordinary (later ordinary) professor at the University of Berlin. In 1831 he became a member of the Berlin Academy of Sciences, and in the same year he married Rebecca Mendelssohn-Bartholdy, sister of Felix Mendelssohn,

the composer. Dirichlet spent twenty-sevenyears as a professor in Berlin and exerted a strong influence on the development of German mathematics through his lectures, through his many pupils, and through a series of scientific papers of the highest quality that he published during this period. He was an excellent teacher, always expressing himself with great clarity. His manner was modest; in his later years he was shy and at times reserved. He seldom spoke at meetings and was reluctant to make public appearances. In many ways he was a direct contrast to his lifelong friend, the mathematician Karl Gustav Jacobi.

One of Dirichlet's most important papers, published in 1850,deals with the boundary value problem, now known as Dirichlet's boundary value problem, in which one wishes to determine a potential function satisfying Laplace's equation and having prescribed values on a given surface, in Dirichlet's case a sphere. In 1855, when Gauss died, the University of GOttingen was anxious to seek a successor of great distinction, and the choice fell upon Dirichlet. Dirichlet moved to Gottingen in the fall of 1855, bought a house with a garden, and seemed to enjoy the quieter life of a prominent university in a small city. He had a number of excellent pupils and relished the increased leisure for research. His work in this period was centered on general problems of mechanics. This new life, however, was not to last long. In the summer of 1858 Dirichlet traveled to a meeting in Montreux, Switzerland, to deliver a memorial speech in honor of Gauss. While there, he suffered a heart attack and was barely able to return to his family

in Giittingen. During his illness his wife died of a stroke, and Dirichlet himself died the followingspring.

Some special cases of (22.2) are worthy of mention. The first is U(Xb) = 0, the solntion to an inhomogeneous DE satisfying the homogeneous BC. We obtain this by substituting zero for g(Xb) in (22.2) so that only the integration over D remains. The second special case is when the DE is homogeneous, that is, when f(x) = but the BC is inhomogeneous. This yields an integration over the bonndary alone. Finally, the solution to the homogeneous DE with the homogeneous BC is simply U = 0, referred to as the zero solution. This is consistent with physical intuition: If the function is zero on the boundary and there is no source f (x) to produce any "disturbance," we expect no nontrivial solution.

°

aD

22.1.1. Example. Let us lind the Green's function for the three-dimeosional Laplacian Lx = V2 satisfyingthe Dirichlet Be G D(P, y) = 0 for P, on the xy-plane. Here D is the upper half-space (z '" 0) and aD is the xy-plane. It is more convenient to use r = (x, y, z) and r' = (x', s', z') instead of x and y, respectively. Using (21.21) as Gg), we can write , I , GD(r,r)=-4 I 'I +H(r,r) rrr-r

616

22. MULTIDIMENSIDNAL GREEN'S FUNCTIDNS: APPLICATIDNS

I

I

- - 411" )(x _ x')2

+ (y -

yl)2

+ (z

Z') 2

+ H(x, y, z: x', v', Zl).

The requirement that G D vanish on the xy-plane gives

,

"

H(x,y,O;x ,y ,z) =

1

1

411" )(x _ x ')2

+ (y

_ yl)2

+ zrJ.

This fixes the dependence of H on all variables except z. On the other hand, V 2 H = 0 in

D implies that the form of H must be the same as that of Gg) because except at r =

t',

Gg)

the latter does satisfy Laplace's equation. Thus, because of the symmetry of in r and r' [GD(r, r') = GD(r', r)] and the evenness of the Laplacian in z (as well as x and y), we have two choices for the z-dependence: (z - z')2 and (z + z')2. The first gives G D = 0, which is a trivial solution. Thus, we must choose

--;===;,,;;==;,=I~7"F=:===;'7i

H(x,y,z;x I ,y " ,z) = 1 411" .,j(x _ x l)2

+ (y -

yl)2

+ (z + z')2

=

Note that with r" ee (x', s'. -z'), this equation satisfies \72H -li(r - r"), and it may appear that H does not satisfy the homogeneous DE, as it should. However, r" is outside D, andr i= r" as long as rED. So H does satisfy the homogeneous DE in D. The Green's function for the given Dirichlet Be is therefore

,1(1 I) Ir-r'I-lr-r"l '

GD(r,r)=-411"

method of images

where r" is the reflection of r' in the xy-plane. This result hasa direct physical interpretation. Ifdetermining the solution of the Laplace equation is considered a problem in electrostatics, then Gg>(r, r') is simply the potential at r of a unit point charge located at r', and GD(r, r') is the potential of two point charges of opposite signs, one at r' and the other at the mirror image of r', The fact that the two charges are equidistant from the xy-plane ensures the vanishing of the potential in that plane. The introduction of image charges to ensure the vanishing of G D at 8 D is common in electrostatics and is known as the methodof images. This method reduces the Dirichlet problem for the Laplacian to finding appropriate point charges outside D that guarantee the vanishing of the potential on aD. For simple geometries, such as the one discussed in this example, determination of the magnitudes and locations of such image charges is easy, rendering the method extremely useful. Having found the Green's function, we can pose the general Dirichlet BVP: V 2u = -per) and u(x, y, 0) = g(x, y) for z > O. The solution is u(r) =

4~ +

i: i: r dx'

1""

-00

dy'

dz' per') (Ir

dxll"" dy'g(x'y') 8GD -00

~ r'1 - Ir ~ r'11)

(22.3)

I '

8z z=o

where r = (x, y, z), r' = (x', s'. z'), and r' = (x', v', -z'). A typical application consists in introducing a number of charges in the vicinity of an infinite conducting sheet, which is held at a constant potential Vo. If there are N charges,

22.1 ELLIPTIC EQUATIONS

617

{qil~l' located at{ril~l' then p(r) = I:~I qj8(r - ri), g(x, y) = const = Va, and we gel

~

u(r) = N - 1 4rr

qi) 1 '1 r~

(qi

00

----[r - ri I [r I

+Vo

00

dx

-00

-00

dy ,80D -8z

Iz=o'

(22.4)

=

where ri (Xi, Yi, Zi) and ~ = (Xi, Yi, -Zi)· That the double integral in Equation (22.4) is unity can be seen by direct integration or by noting that the sum vanishes when z = o. On the other hand, u(x, y, 0) = Vo- Thus, the solution becomes u(r) =

t ~ (_q_i- - ----'!!...-)

i=1 4".

Ir-ril

+ Va.

III

Ir-ril

22.1.2. Example. The method of images is also applicable whenthe boondary is a sphere. Inside a sphere of radius a with center at the origin, we wish to solve this Dirichlet BVP: V 2u = -p(r, e, rp) for r < a, and u(a, e, rp) = g(e, rp). The GF satisfies

v2a otr. e, rp; r', o', rp') = 8(r GD(a, e, ip; r', «, cp') = o.

r')

for r < a,

(22.5)

Thus, G D can again be interpreted as the potential of point charges, of which one is in the sphere and the others are outside.

We write GD = G~)

+H

and choose H in such a way that the second equation in

(22.5) issatisfied. As in the case of the xj-plane, let l H (r, r") =

k ,,' where k is 4".lr-r I a constant to be determined. If r" is outside the sphere, V 2 H will vanish everywhere inside the sphere. The problem has been reduced to finding k and r" (the location of the image

charge). We want to choose r" such that -I I 'I

I

rrr=a

I

= I _k r"1 r r=a

=}

'

"

k(lr - r I)r= = (Ir - r I)r~a.

This shows that k must be positive. Squaring both sides and expanding the result yields

kZ(aZ + r'Z _ 2ar' cosy)

= aZ + r"Z -

2ar" cos y,

where y is the angle between r and r', and we have assumed that r' and r" are in the same direction. If this equation is to hold for arbitrary y, we must have kZr' = r'' and k 2(a2 +r,2) = a 2 +r,l2. Combining these two eqnations yields k 4r12 _k 2 (a 2+r12)+a2 = 0, whose positive solutions are k = 1 and k = aI r. The first choice implies that r" = r', which is impossible because r" must be outside the sphere. We thus choose k = air', which gives r" = (a 2Ir'Z)r'. We then have

, 1[ 1

ar'] . GD(r,r)=-- - - 4". [r - r'l Ir,2r - a 2r'I

(22.6)

1Actually, to be general, we must add an arbitrary function !(r") to this. However, as the reader can easily verify, the following argument will show that! (r") = O. Besides, we are only interested in a solution, not the most general one. All simplifying assumptions that follow are made for the same reason.

618

22. MULTIDIMENSIDNAL GREEN'S FUNCTIDNS: APPLICATIDNS

Substituting this in Equation (222), audnoting that aG/any = (aG/ar')"=a, yields u(r) =

~ 411"

{a rfldr'

Jo

+ a(a 2 -

r

2)

t" sinO' so'

10

r:

-

(~r'1 [r -

2 ar' 2

IT' r - a

r'l

) p(r') dq/

dq/ (" siu 0' dO' g(O',q/) , [r - al 3

10

4"

{2"

10

(22.7)

10

where a = (a, 0', q/) is a vector from the origin to a pointon the sphere. FortheLaplace equation p(r') = 0, aud only the double integral in Equation (22.7) will contribute. It cau be shown that if g(O', '1") = const = Vo, then u(r) = Vo. This is the familiar factshown in electromagnetism: If thepotential on a sphere is kept constant, thepotential inside the sphere will be constantand equal to the potential at the surface. II

22.1.3. Example. In thisexample we findthe Dirichlet GF for a circle of radius a centered at the origin. The GF is logatithutic [seeEquation (21.22)]. Therefore, H is also logatithutic, and its mostgeneral form is H(r, r") =

-~ In(lr 2"

r"l) -

~ in[f(r")] = -~ In(lr 2"

2n

r"lf(r")),

so that

,1 ,1 " "))=-In 1 1 r-r' G D(r,r)=-In(lr-rl)--In(lr-r If(r r" " 2" 2" 2n (r )f(r)

1

ForGD to vanish atall points on the circle, we musthave

3-r'

I

(a _ r")f(r") = I

=}

"

,

[a - r'[ = I(r - r )f(r' )1,

1

where a is a vector from origin to a pointon the circle. Assuming that r" andr' are in the same direction, squaring bothsidesof thelastequation andexpanding theresult, we obtain (a 2 + r '12 - 2ar" cos y) f 2 (r") = a Z + r t2 - 2ar' cos y, where y is theanglebetween a and r' (orr"). This equation must hold forarbitrary y. Hence, we have j2(r")r" = r' and j2(r")(a 2 + r,12) = a 2 + r f2. Thesecanbe solvedfor j(r") andr", Theresult is 2

r',=a r, ,2 '

"a r' r = -. a Substituting theseformulas in theexpression for GD, we obtain r

f(r ) = "

(I a '1r') 2

, I , I G D(r,r)=-In(lr-rl)--1n 23f 23f

r--r r 12 a

.

To write the solution to the Dirichlet BVP, we also need aGD/an = aGD/ar'. Using polar coordinates, we express GD as

I

I

, I r 2 + r,2 - 2rr' cos(t9 - 19') G D(r,r)=-1n . 2,12 4" r /a 2 + a 2 - 2rr' cos(O - 0')

Differentiation withrespect to r' yields

aGDI aGD\ ---a;- r'=a = a;:r r'=a =

2_r2

I a 23fa r 2 + a 2 - 2ra cos(t9 - 19')'

22.1 ELLIPTIC EQUATIONS

619

from whichwe canimmediately write the solution to thetwo-dimensional Dirichlet BVP V 2u = p, u(r = a) = g(e') as u(r) =

Poisson integral formula

2n

foo

de'

loa 0

+ -a 2- -r- fo2n de' 2

r'GD(r,r')p(r')dr'

2

"a

0

2

g(e')

2

r +a -2racos(e-e')

.

In particular, forLaplace's equation p(r') = 0, and we get 2 2 u(r, e) = a - r

2"a

r:

de'

10

g(e')

r 2 + a2 - 2ra cos(e - 0')

.

(22.8)

Equation (22.8)is calledthe Poisson integral formula.

22.1.2

11II

The Neumann Boundary Value Problem

The Neumann BVP is not as simple as the Dirichlet BVP because it requires the normal derivative of the solution. But the normal derivative is related to the Laplacian through the divergence theorem. Thus, the BC and the DE are tied together, and unless we impose some solvability conditions, we may bave no solution at all. These points are illustrated clearly if we consider the Laplacian operator. Consider the Neumann BVP

v2u =

f(x)

for xED,

au an

and

= g(x)

for x

E

aD.

Integrating the first equation over D and using the divergence theorem, we obtain

It follows that we cannot arbitrarily assign values of au/an on the boundary. In particular, if the BC is homogeneous, as in the case of Green's fuuctions, the RHS is zero, and we must have f(x) dmx = O. This relation is a restriction on the DE, and is a solvability condition, as mentioned above. To satisfy this condition, it is necessary to subtract from the inhomogeneous term its average value over the region D. Thus, if VD is the volume of the region D, then

JD

f- = - I

where

VD

J

f(x)dmx

D

ensures that the Neumann BVP is solvable. In particnlar, the inhomogeneous term for the Green's function is not simply 8(x - y) but 8(x - y) - g, where

-= -11

8

VD D

8(x - y) dm x

= -I

VD

if y E D.

620

22. MULTIDIMENSIONAL GREEN'S FUNCTIONS: APPLICATIONS

Thus, the Green's function for the Neumann BVP, GN(X, y), satisfies I V 2GN(X, y) = 8(x - y) - - ,

Vn

aGN --(x, y)

an

=a

for x E aD.

Applying Green's identity, Equation (21.27), we get u(x) = ( dmyGN(X, y)f(y) -

In

(

laD

GN(X, y) aau da

n

+ ii,

(22.9)

where ii = (l/Vn) In u(x) dmx is the average value of u in D. Equation (22.9) is valid only for the Laplacian operator, although a similar result can be obtained for a general self-adjoint SOLPDO with constant coefficients. We will not pursue that result, however, since it is of little practical use.

Carl Gottfried Neumann (1832-1925) was the son of Franz ErnstNeumann, a professorof physics andmineralogyat Konigsberg; his mother, Luise Florentine Hagen, was a sisterin-law of theastronomer Bessel. Neumann receivedhis primary andsecondaryeducationin Konigsberg, attended theuniversity, and formed particularly close friendships with the analyst F. J. Richalot andthegeometerL.a.Hesse. Afterpassingtheexaminationforsecondary-school teaching, he obtained his doctorate in 1855; in 1858 he qualifiedforiectnring in mathematics at Halle, where he became Privatdozent and, in 1663,assistant professor. In the latter year he was called to Basel, and in 1865 to Tiibingen. From the autumn of 1868 untii his retirement in 1911 he was at the University of Leipzig. In 1864 he married Hermine Mathilde Elise Kloss; she diedin 1875.

Neumann, who led a quietlife, was a successful university teacher anda productive researcher. More than two generations of future gymnasium teachers received their basic mathematical education from him. As a researcher he wasespecially prominent in thefield of potential theory. His investigations into boundaryvalue problems resulted inpioneering achievements; in 1870 he began to develop themethod of the arithmetical mean for their solution. He alsocoinedtheterm "logarithmic potential." Thesecondboundary value problem of potential theory stillbears his name; a generalization of it was later provided by H. Poincare. Neumann wasamember of the BerlinAcademy andtheSocietiesofGoningen, Munich, and Leipzig. Heperformed avaluable service infounding and editing theimportant German mathematics periodical Mathematische Annalen.

Throughout the discussion so far we have assumed that D is bounded; that is, we have considered points inside D with Bes on the boundary D specified. This is called an interior BVP. In many physical situations we are interested in points

a

interior vs exterior

BVP

22.2 PARABOLIC EQUATIONS

621

outside D. We are then dealing with an exterior BVP. In dealing with such a problem, we must specify the behavior of the Green's function at infinity. In most cases, the physics of the problem dictates such behavior. For instance, for the case of an exterior Dirichlet BVP, where u(x) = .

1

d"'yGD(X, y)f(y)

D

+

fa

aD

aGD U(Yb)--(X, Yb) da any

and it is desired that u(x) --> 0 as [x] --> 00, the vanishing of GD(X, y) at infinity guarantees that the second integral vanishes, as long as aD is a finite hypersurface. To guarantee the disappearance of the first integral, we must demand that G D (x, y) tend to zero faster than f(y)d"'y tends to infinity. For most cases of physical interest, the calculation of the exterior Green's functions is not conceptually different from that of the interior ones. However, the algebra may be more involved. Later we will develop general methods for finding the Green's functions for certain partial differential operators that satisfy appropriate BCs. At this point, let us simply mention what are called mixed BCs for elliptic POEs. A general mixed BC is of the form a(x)u(x)

au + fl(x)-(x) =

an

y(x).

(22.10)

Problem 22.6 examines the conditions that the GF must satisfy in such a case.

22.2 Parabolic Equations Elliptic partial differential equations arise in static problems, where the solution is independent of time. Of the two major time-dependent equations, the wave equation and the heat (or diffusion) equation.e the latter is a parabolic POE and the former a hyperbolic POE. This section examines the heat equation, which is of the form "12u = a 2au/at. By changing t to t/a 2 , we can write the equation as Lx,,[u] sa (a/at - "12)u(x, r) = 0, We wish to calculate the Green's function associated with Lx" and the homogeneous BCs. Because of the time variable, we must also specify the solution at t = O. Thus, we consider the BVP

Lx ,[u] es (~ - "12) u(x, r) = 0 , at U(Xb, t)

= 0,

u(x,O)

= hex)

for xED, for Xb E aD, xED.

(22.11)

To find a solution to (22.11), we can use a method that turns out to be useful for evaluating Green's functions in general-the method of eigenfunctions. Let 2The heat equation turns into the Schrcdinger equation if t is changed to the Schrcdlnger equation as well.

A

t; thus, the following discussion incorporates

622

22. MULTIDIMENSIONAL GREEN'S FUNCTIONS: APPLICATIONS

[un}~1 be the eigenfunctions of '1 2 with eigenvalues [-An}~I' Lei the Un(Xb) = 0 for Xb E an. Then '1 2Un (X)

+ Anun(x) =

0 Un(Xb) = 0

Be be

for n = 1,2, ... , xED, for Xb E an.

(22.12)

Equation (22.12) constitutes a Sturm-Liouville problem io m dimensions, which we assume to have a solution with [un}~1 as a complete orthonormal set. We can therefore write 00

u(x, I) =

L Cn(/)un(x).

(22.13)

n=l

This is possible because at each specific value of I, u(x, I) is a function of x and therefore can be written as a lioear combioation of the sameset, [un}~l' The coefficients Cn (I) are given by Cn(1) =

In

u(x, I) Un (x) d"»,

(22.14)

To calculate Cn(/), we differentiate (22.14) with respect to time and use (22.11) to obtaio (;n(t)

==

ddCn = [ au (x, I)Un (x) dmx = [ ['1 2 u (x, I)]un(x)dmx. I JD al JD

Usiog Green's identity for the operator '1 2 yields [ [un'1 2u _ u'1 2un] dmx = [ (Un au _ U aUn) da. JD Jan an an

Sioce both U and Un vanish on an, the RHS is zero, and we get (;n(t) =

In

u'12u n dmx = -An

In

u(x, I)Un(X) dmx = -AnCn.

This has the solution Cn(/) = Cn(O)e-A"t, where Cn(O)

=

In

u(y, O)un(y)dmy

=

In

h(y)un(y)dmy,

so that Cn(t) = e- Ant

In

h(y)un(y) dmy.

Substitutiog this io (22.13) and switching the order of integration and summation, we get u(X, I) =

1 D

[f:e-A"tun(X)Un(y)] h(y)dmy n=l

22.2 PARABOLIC EQUATIONS

623

and read off the GF as L~! e-J.,IUn(x)Un(y)e(I), where we also introduced the theta function to ensure that the solution vanishes for I < O. More generally, we have 00

G(x, y; I - r) =

L e-J."CI-')Un(x)un(y)e (I -

r).

(22.15)

n=l

Note the property 00

lim G(x, y; I - r)

7:-+1

= "un(x)un(y) = 8(x L.."

y),

n=l

which is usually written as G(x, y; 0+) = 8(x - y).

(22.16)

The reader may also check that LX,IG(X, y; I - r). = 8(x - y)8(1 - r)

(22.17)

This is precisely what we expect for the Green's function of an operator in the variables x and I. Another property of G(x, y; I - r) is that it vanishes on aD, as it should. Having found the Green's function and noted its properties, we are in a position to solve the inhomogeneous analogue of Equation (22.11), in which the RHS of the first equation is f (x, I), and the zero on the RHS of the second equation is replaced by g(Xf" I). Experience with similar but simpler problems indicates that to make any progress toward a solution, we must come up with a form of Green's identity involving Lx,1 and its adjoint. It is easy to show that vLx,I[U]- U~,,[V] =

a

al (uv)

- V . (vVu - uVv),

(22.18)

where LIt = -a/al - v 2 • Now consider the (m + I)-dimensional "cylinder" one of whose bases is at I = E, where E is a small positive nmnber. This base is barely above the mdimensional hyperplane IR m . The other base is at I = r - E and is a duplicate of D C IRm (see Figure 22.1). Leta", where", = 0, 1, ... , m, be the components of an (m + I)-dimensional vector a = (ao, ai, ... , am). Define an inner product by m

a· b es

L a"b" sa aObo - alb! -

... - amb m as aObo - a- b

p,=O

and the (m + I)-dimensional vector II by QO (22.18) can be expressed as

m aQ"

aQo

= UV, Q = vVu aQI

aQm

vLxt[u] - uL t [v] = , , - - '" - - - - - - ... - - -m. , X,I L.. ax" ax o axl ax ,,~o

uVv. Then

(22.19)

624

22. MULTIDIMENSIDNAL GREEN'S FUNCTIDNS: APPLICATIDNS

Figure 22.1 The "cylinder" used in evaluating the GFfor thediffusion and waveequations. Note that thebasesarenotplanes, buthyperplanes (that is, spaces suchas JRrn).

We recognize the RHS as a divergence in (m + I)-dimensional space. Denoting the volume of the (m + I)-dimensional cylinder by D and its boundary by a:D and integrating (22.19) over D, we obtain

(22.20) where dS is an element of "area" of a'D. Note that the divergence theorem was used in the last step. The LHS is an integration over t and x, which can be written as

The RHS of (22.20), on the other hand, can be split into three parts: a base at a base at t = 'r - E, and the lateral surface. The base at t = E is simply the region D, whose outward-pointing normal is in the negative t direction. Thus, no = -1, and n, = 0 for i = 1,2, ... , m. The base at t = r - E is also the region D; however, its DOnna! is in the positive t direction. Thus, no = 1, and n; = 0 for i = 1,2, ... , m. The element of "area" for these two bases is simply dmx. The unit normal to the lateral surface has no time component and is simply the unit normal to the boundary of D. The element of "area" for the lateral surface is dt da, where da is an element of "area" for aD. Putting everything together, we t

= E,

22.2 PARABOLIC EQUATIONS

625

can write (22.20) as

1'-' in dt

dmx(vLx.'[u] - uLI., [v])

= {(_QO)I,=,dmx+ {

In

In

QOI,~,_,dmx-

(

Jan

da1'-'dtQ.en. e

The minus sign for the last term is due to the definitiou of the inner product. Substituting for Q yields

1'-' in -in dt

dmx(vLx,,[u] - uLI,,[v])

u(x, f)V(X, f)dmx

=

-

fa

da

,-, 1

+

in

u(x, r - f)V(X, r - f)dmx (22.21)

(au av) dt v - - u - . an an

aD,

Let v be g(x, y; t - r), the GF associated with the adjoint operator. Theu Equation (22.21) gives

1'-' in -in ( 1'-' [ - laD dt

dmx[g(x, y; t - r)f(x, t) - u(x, t)8(x - y)8(t - r)] u(x, f)g(X, y; f - r)dmx

=

da,

+

in

u(x, r - f)g(X, y; -f)dmx

g] a . dt g(Xb, y; t - r) au an - U(Xb, t) an (22.22)

We uow use the following facts: 1. 8(t - r) = 0 in the second integral on the LHS of Equation (22.22), because t can never be equal to r in the range of integration. 2. Using the symmetry property of the Green's function and the fact that Lx" isreal, we have g(x, y; t - r) = G(y, x; r -t), where we have used the fact that t and r are the time components of x and y, respectively. In particular, by (22.16), g(x, y; -f) = G(y, x; f) = 8(x - y). 3. Thefunctiong(x, y; t-r) satisfies the sarne homogeneous BC as G(x, y; tr). Thus, g(Xb, y; t - r) = 0 for x, E aD. Substituting all the above in (22.22), taking the limit f --> 0, and switching x and y and t and r, we obtain u(x, t) =

fa' dr in dmyG(x, y; t - r)f(y, r) + in u(Y, O)G(x, y; t)dmy -

l°'

dr

fa et»

aG Yb; t - r) da, U(Yb, r)-(x, any

(22.23)

626

22. MULTIDIMENSIONAL GREEN'S FUNCTIONS: APPLICATIONS

where a/any in the last integral means normal differentiation with respect to the second argument of the Green's function. Equation (22.23) gives the complete solution to the BVP associated with a parabolic PDE. If f (Y, r) = 0 and U vanishes on the hypersurface aD, then Equation (22.23) gives u(x, t) = GF as evolution

operator or propagator

L

(22.24)

u(y, O)G(x, Y; t)dmy,

which is the solution to the BVP of Equation (22.11), which led to the general Green's function of (22.15). Equation (22.24) lends itself nicely to a physical interpretation. The RHS can be thought of as an integral operator with kemel G(x, y; t). This integral operator acts on u(y, 0) and gives u(x, t); that is, given the shape of the solution at t = 0, the integral operator prodnces the shape for all subsequent time. That is why G(x, y; t) is called the evolution operator, or propagator.

22.3 Hyperbolic Equations The hyperbolic equation we will discuss is the wave equation Lx.t[u] '"

(::2 - '1,72) u(x, t)

(22.25)

= 0,

where we have set the speed of the wave equal to unity. We wish to calculate the Green's function for Lx,t subject to appropriate Bes. Let us proceed as we did for the parabolic equation and write 00

G(x, y; t) =

L Cn(y; t)un(x)

Cn(y; t) =

n=l

1 D

G(x, y; t)un(x) dmx, (22.26)

where Un (x) are orthonormal eigenfunctions of '1,72 with eigenvalues -An, satisfying certain, as yet unspecified, Bes. As usual, we expect G to satisfy

. = (aat

2

Lx,t[Gj

2

-

'1,72 )

G(x, y; t - r)

Substituting (22.26) in (22.27) with r

= 8(x -

y)8(t - r),

(22.27)

= 0 and using 'I,72 un = -Anun, gives

where we used 8(x - y) = L::'1 Un (x)Un(y) on the RHS. The orthonormality of Un now gives tn(y; t) + AnCn(y; t) = un(y)8(t). It follows that Cn(y; t) is

22.3 HYPERBOLIC EQUATIONS

627

separable. In fact, C,,(y; t) = u,,(y)T,,(t)

where

(:t:

+ An) Tn(t) =

8(t).

This equation describes a one-dimensional Green's function and can be solved using the methods of Chapter 20. Assuming that Tn(t) = 0 for t :'" 0, we obtain T,,(t) = (sinw"t/w,,)e(t), where w~ = An. Substituting all the above results in (22.26), we obtain ~ sinwnt G(x, y; t) = L..,un(x)un(y)--O(t), CUn

n=l Of,

more generally, ~

sinwn(t - r)

n=l

liJn

G(x, y; t - r) = L..,un(x)un(y)

O(t - c).

(2228)

We note that G(x, y; 0+) = 0

eo at

-(x, y; t) I

and

= 8(x - y),

(22.29)

t--).O+

as can easily be verified. With the Green's function for the operator Lx,1 of Equation (22.25) at our disposal, we can attack the BVP given by

(:t

22

U(Xb, t)

2

- '1 )

= h(Xb, r),

u(x,

t)

u(x,O)

= !(x, r)

for xED,

= t/J(x)

for

U

aa (x, t) I = 1/r(x) t t=O

Xb

E aD,

xED,

for xED.

(22.30)

As in the case of the parabolic equation, we first derive an appropriate expression of Green's identity. This can be done by noting that

vL flu] x,

U"

t [v] = -a (av u - - vau) - - V . (uVv - vVu).

-X,I

at

at

at

Thus, Lx,1 is formally self-adjoint. Furthermore, we can identify

o

av at

au at

Q =u--v-

and

Q= uVv - vVu.

628

22. MULTIDIMENSIONAL GREEN'S FUNCTIONS: APPLICATIONS

Following the procedure nsed for the parabolic case step by step, we can easily derive a Green's identity and show that u(x, t) =

fot dt: 10 dmyG(x, y; t (

+ 1D [1ft(y)G(x, y; t) -

-1r 0

dx

r

laD

7:)f(y, 7:)

aG

(Y)iJt(x, Y; t)]dmy

h(Yb, 7:) aG (x, Yb; t - 7:) da.

any

The details are left as Problem 22.11. For the homogeneons PDE with the homogeneous BC h = 0 = u(x, r) = -

f

D

(22.31)

1ft, we get

aG (y)-(x, Y; t) d m y.

at

Note the difference between this equation and Equation (22.24). Here the propagator is the time derivative of the Green's function. There is another difference between hyperbolic and parabolic equations. When the solution to a parabolic equation vartishes on the boundary and is initially zero, and the PDE is homogeneous [f(x, t) = 0], the solution must be zero. This is clear from Equation (22.23). On the other hand, Equation (22.31) indicates that under the same circumstance, there may be a nonzero solution for a hyperbolic equation if 1ft is nonzero. 10 such a case we obtain u(x, r) =

10 1ft (y)G (x, y; t)dmy.

This difference in the two types of equations is due to the fact that hyperbolic equations have second-order time derivatives. Thus, the initial shape of a solution is not enough to uniquely specify it. The initial velocity profile is also essential. We saw examples of this in Chapter 19. The discussion of Green's functions has so far been formal, The main purpose of the remaining sections is to bridge the gap between founalism and concrete applications. Several powerful techniques are used in obtaining Green's functions, hnt we will focus only on two: the Foutier transfoun technique, and the eigenfunction expansion technique.

22.4

The Fourier Transform Technique

Recall that any Green's function can be written as a sum of a singo1ar part and a regular part: G = G, + H. Since we have already discussed homogeneous equations in detail in Chapter 19, we will not evaluate H in this section but will concentrate on the singular parts of various Green'8 functions. The BCs play no role in evaluating G,. Therefore, the Foutier transfoun technique (FTT), which involves integration over all space, can be utilized. The FTT

22.4 THE FOURIER TRANSFORM TECHNIQUE

629

has a drawback-it does not work if the coefficient functions are not constants. For most physical applications treated in this book, however, this will not be a shortcoming. Let us consider the most general SOLPDO with constant coefficients,

a

m

Lx = ao + ~:>Fa-: )=1

XJ

a2

m

+

L

j,k=l

bjkaT'

(22.32)

X J Xk

where ao, aj, and bjk are constants. The corresponding Green's function has a singular part that satisfies the usual PDE with the delta function on the RHS. The FIT starts with assuming a Fourier integral representation in the variable x for the singular part and for the delta function: G,(x,y) =

I /2 jdmkG,(k,y)e ik ' X ,

(21f)m

8 (x - y) = _1_ (21f)m

jd

m keik.(x-y).

Substituting these equations in the PDE for the GF, we get

and

(22.33)

If we can evaluate the integral in (22.33), we can find G. The following examples apply Equation (22.33) to specific problems. Note that (22.33) indicates that G, depends only on x - y. This point was mentioned in Chapter 20, where it was noted that such dependence occurs when the BCs play no part in an evaluation of the singular part of the Green's function of a DE with constant coefficients; and this is exactly the situation here.

22.4.1

GF for the m-Dimensional Laplacian

We calculated the GF for the m-dirnensional Laplacian in Section 21.2.2 using a different method. With ao = 0 = aj, bjl = 8jz, and r = x - y, Equation (22.33)

reduces to I G,(r) = (21f)m

j

eik .• dmk _k2'

(22.34)

where k 2 = kr + ... + k~ = k . k. To integrate (22.34), we choose spherical coordinates in the m-dirnensional k-space. Furthermore, to simplify calculations

630

22. MU l TI DIMENS IONAl GREEN'S 'FUNCTIONS: APPll CATIONS

we let the km-axis lie along r so that r = (0,0, ... , [r[) and k . r = klrl cos{h [see Equation (21.12)]. Substitutiog this in (22.34) and writiog dmk in spherical coordinates yields

From Equation (21.15) we note that drl.m = (sinet)m-2deldrl.m_l. Thus, after em-t, Equation (22.35) becomes integrating over the angles

ez, ... ,

G (r) = __I_r1. _ s

(21f)m

m 1

10(00 k

m-3dk ("(sine )m-2eiklrlcoSOlde .

10

t

1

The inner integral can be looked up in an integral table (see [Grad 65, p. 482]):

10r

=.fii

(sinet)m-2eiklrlcosOldel

(kr2)m/2-t f (m--T-1) Jm/2- 1(k r ).

Substitutiog this and (21.16) in the preceding equation and using the result (see [Grad 65, p. 684])

we obtain f(m/2 -

4nrn/2

1) (_1_)

for m > 2,

r m-2

which agrees with (21.20) since f(m/2) = (m/2 - 1)r(m/2 - I).

22.4.2

GF for the m-Dimensional Helmholtz Operator

For the Hehnholtz operator '17 2 1 Gs(r) = - (21f)m

f'

- J.t2,

e

dmk J.t2

ik

'

Equation (22.33) reduces to

r

+ k 2'

Following the same procedure as in the previous subsection, we find

22.4 THE FOURIER TRANSFORM TECHNIQUE

"631

Here we can use the integral formula (see [Grad 65, pp. 686 and 952])

where

to obtain

rrr -

Q (2)m/2-1 G (r) = -~ s (2:n:)""'" r

- I) n m/2-1_e,mn/4H(I) :n: . r (m --" 2""

2

(inr) m/2-1 "" ,

which simplifies to

:n:/2 (fh)m/2-1 imn/4 (I) . G,(r) = - (2:n:)m/2 -; e Hm/2- 1('W)· .

It can be shown (see Problem 22.8) that for m

(22.36) e-/l- r

= 3 this reduces to G,(r) = - -:n:r 4'

which is the Yukawa potential due to a unit charge. We can easily obtain the GF for \72 +fh2 by substituting ±ifh for fh in Equation (22.36). The result is G,(r)

:n:/2 (fh)m/2-1 =,.m+! (2:n:)m /2 H m / 2 1(±W)· r (I)

(22.37)

For m = 3 this yields G,(r) = _e±i/U/(4:n:r). The two signs in the exponent correspond to the so-called incoming and outgoing "waves." Non-local potentials depend not only on the observation point, but also on some other "non-local" variables.

22.4.1. Example. For a non-localpotential, the time-independentSchriidingerequation is

Then, the integral equation associated withthis differential equation is (see Section 21.4) \II(r)

= Aeik or

_

---.!:!:..- {

2rr1i2

JR'

d 3r, i

k1r - r'1 (

[r - r']

JR'

d 3r"V(r', r")lJ1(r").

(22.38)

Fora separablepotential,for which V(r ', r") = -g2U(r')U(r"), we can solve(22.38) exactly. We substitutefor V(r ', r") in (22.38) 10 obtain (22.39)

632

22. MULTIOIMENSIONAL GREEN'S FUNCTIONS: APPLICATIONS

(22.40)

and substituting themin (22.39) yields 'I1(r) = Aei k·r + C Q(r). Multiplyingboth sidesof this equationby U (r) and integrating over 1R3, we get C= A [

JR3

e'k.rU(r)d3r

+C

[

JR3

U(r)Q(r)d 3r

U(r)Q(r)d 3r,

= (2".)3/2Ail(-k)+C [ lJW.3

from wbich we obtain C = (2Jl")3/2Ail(-k)/[1 - JJW.3 U(r)Q(r)d 3r], leading to the solution

'I1(r) _ Aeik.r + -

3/2 (2".) AU(-k) Q(r) I-JJW.3U(r l)Q(rl)d 3r' ·

(22.41)

In ptinciple, U(-k) [the Fourier transform of U(r)] and Q(r) can be calculatedonce the functional form of U(r) is known. Equations (22.40) and (22.41) give the solution to the Schriidinger equation in closedform. IlII!I

When dealing with parabolic and hyperbolic equations, we will find it convenient to consider the "different" variable (usually r) as the zeroth coordinate. In the Foutier transfonu we then use co = -ko and write

1 ! 00

Gs ( r, ) t

1 = (2".) (m+l)J2

8(r)8(t) =

I (2". )m+t

.1

d (J)

-00

dmkG- s(k, w ) i(k·r-wt) e.

00

dW!dmkeirJ<.r-wt)

-00

(22.42) '

where r is the m-dimensional positionvector.

22.4.3

GF for the m-Dimensional Diffusion Operator

We substitote from (22.42) in (a/at - V 2)G,(r, t) = 8(r)8(t) to obtain G,(r, t) =

1 +1

(21f)m

! .1

iwt dco e- . 2' -00 W + ik 00

dmke,k.r

I:;:1

(22.43)

kr. The to integration can be done using the calculus where as usual, k 2 = of residues. The integrand has a simple pole at co = -ik2 , that is, in the lower half of the complex w-plane (LHP). To integrate, we must know the sign of t. If t > 0, the exponential factor dictates that the contour be closed in the LHP, where there is a pole and, therefore, a contribution to the residues. On the other hand, if t < 0, the contour must be closed in the UHP. The integral is then zero because

633

22.4 THE FOURIER TRANSFORM TECHNIQUE

there are no poles in the upper half plane (UHF). We must therefore introduce a step function 0 (r) in the Green's function. Evaluating the residue, the '" integration yields - 2rr i e-k " . (The minus sign arises because ofclockwise contour integration in the LHP.) Substitnting this in Equation (22.43), using spherical coordinates in which the last k-axis is along r, and integrating over all angles except 01, we obtain Q

G,(r, r) = 8(t) (2:)~

1roo km-ldke- k'a 1t" (sinOl)m-2eikrcosOldOl 0

0

=O(t) m-l..;rr (2)m/2-l -r r (m -2-I) 1n0 (2n)ffl Q

00

a

k m/2e-k 'Jm/2_ l (kr) d k .

For the 01 integration, we used the result quoted in Section 22.4.1. Using the integral formula (see [Grad 65, pp. 716 and 1058])

roo I' _ax' 10 x e Jv(fJx)dx =

v

fJ r

(!L+ V+ 1)

2 2 v+ lct(I'+v+l)/2r(v + I)

(!L+V+1·. 2

'

V

+ I, -

fJ2) 4ct '

where is the confluent hypergeometric function, we obtain _ 21l'(m-l)/2 G,(r, t) - 8(t) (21l')m

(2)m/2-l r m/2-l (m m r2 ) ..;rr ;: 2m/2tm/2 2' 2' - 4t

.

(22.44)

The power-series expansion for the confluent hypergeometric function shows that (o, a; z) = e': Substituting this result in (22.44) and simplifying, we finally obtain

o,«. t) = 22.4.4

e- r2/ 4t

(22.45)

( /20(t). 4rrt)m

GF for the m-Dimensional Wave Equation

The difference between this example and the preceding one is that here the time derivative is of second order. Thus, instead of Equation (22.43), we start with G,(r, r) =

I (2rr)m+l

f

d m ke i k .r

1

iwt

00

-00

dt» e",2 -

k 2'

(22.46)

The ci integration can be done using the method of residues. Since the singularities of the integrand, '" = ±k, are on the real axis, it seems reasonable to use the principal value as the value of the integral. This, in turn, depends on the sign of t. If t > 0 (t < 0), we have to close the contour in the LHP (UHF): to avoid the explosion of the exponential. If one also insists on not including the poles inside the conrour.' then one can show that e-iwt

00

P

1

-00

dto

",2 _k2

sin kt = -rr--«t), k

3This will determine how to (semi)circle around thepoles.

634

22. MULTIDIMENSIONAL GREEN'S FUNCTIONS; APPLICATIONS

where E(t) es e(t) _ e(-t) =

{I-1

if t > 0, if t < O.

Substituting this in (22.46) and integrating over all angles as done in the previous examples yields G,(r, r) = Physics determines thecontour of integration

==

0

(22.47)

1

00

e-iwt

dto 2 2 -00 w - k

=

1

e- izt -2--2 dz

c, z - k

2Jr

= -e(t)sinkt. k

The integral is zero for t < 0 because for negative values of t, the contour must be closed in the UHP, where there are no poles inside C I. Substituting this in (22.46) and working through as before, we obtain what is called the retarded Green's function: (ret)

G, advanced Green's funclion

2(2".)"' r"' -

10 00 k",/2-ll",/2-1 (kr ) sm. k t dk .

As Problem 22.25 shows, the Green's function given by Equation (22.47) satisfies only the homogeneous wave equation with no delta function on the RHS. The reason for this is that the principal valne of an integral chooses a specific contour that may not reflect the physical situation. In fact, the Green's function in (22.47) contains two pieces corresponding to the two different contours of integration, and it turos out that the physically interesting Green's functions are obtained, not from the principal value, but from giving small imaginary parts to the poles. Thus, replacing the w integral with a contour iutegral for which the two poles are pushed in the LHP and using the method of residues, we obtain I up

retarded Green's function

E(t) /2 /2 1

e(t) (00 un/2-1 . (r,t) = (2".)m/2 fo K Jm/2_I(kr)smktdk. r",/2-1

(22.48)

If the poles are pushed in the UHP we obtain the advanced Green's function:

10 00 k ",/2-1 J",/2_I(kr) smkt . dk.

e( -t) (2".)"' /2 r'" /2 1 0

(22.49)

Unlike the elliptic and parabolic equations discussed earlier, the integral over k is not a function but a distribution, as will become clear below. To find the retarded and advanced Green's functions, we write the sine term in the integral in terms of exponentials and use the following (see [Grad 65, p. 712]):

(00 fo

v -ax

X e

(2fJ)Vr(v + 1/2) Jv(fJx) dx = ,fii(a2 + fJ2)v+1/2

for Re(a) > 1Im(fJ) I·

635

22.4 THE FOURIER TRANSFORM TECHNIQUE

To ensure convergence at infinity, we add a small negative number to the exponential and define the integral

r:

sa

10

00

kVe-(Tit+,)k Jv(kr) dk =

(2r)Vr~+ 1/2) [('fit + .)2 + r 2r(v+l/2).

For the GFs, we need to evaluate the (common) integral iu (22.48) and (22.49). With v = m/2 - 1, we have I(V) ==

roo kVJv(kr) sinkt dk = 2,2z lim (1: - I;)

io

I

E-?O

(2r)Vr(v + 1/2). = 2i,jii ,~o [r2 + (-it

1

+ <)2]v+I/2

- [r2 + (it

1)

+ <)2]v+1/2

.

At this point, it is couvenient to discuss separately the two cases of m odd and m even. Let us derive the expressiou for odd m (the even case is left for Problem

22.26). Defiue the integer n I(n) =

1)/2 = v +

[1

(2r)n-I/2r(n) lim 2i,jii , ....0 [r2+(_it+.)2]n

Define u = r 2 + (-i t un = (n - 1)1 dun

! and write I(v) as 1 ). [r2 + (it

+ <)2]n

(22.50)

+ .)2. Then using the identity

(_1)n-1 d n -

1

= (m -

I

1 1 (;;-)

and the chain rule, dt/du

1

= (l/2r)af/ar, we obtain d/du = (l/2r)a/ar and 1

[r2 + (±it + .)2]n = (n - 1)1

a )n-I [

(1

- 2r ar

1

]

r 2 + (±it + .)2 .

Therefore, Equatiou (22.50) Canbe written as I(n) =

loo k"-1/2 Jn_ I/2(kr)sinktdk (2r)n-I/2r(n) 1 2i,jii (n 1)1 1 a )n-I [ [ 1 1]) . ( -2rar ,~o [r2+(-it+<)2] - [r2+(it+<)2] .

(22.51)

The limit in (22.51) is found iu Problem 22.27. Using the result of that problem and I'(n) = (n - 1)1, we get I(n) = =

10

00

k n- I/ 2Jn_l/2(kr)sinktdk

,jii(2r)n-I/2 2

(_~!-)n-I [~[8(t + r) _ 2r ar

r

8(t _

-n]. (22.52)

636

22. MULTIDIMENSIONAL GREEN'S FUNCTIONS: APPLICATIONS

Employing this result in (22.48) and (22.49) yields Glret)(r, r) = (adv) G, (r,t) =

Feynman propagator

2.- (__I_~)"-1 [~(t - r)]

ar r 1 ( 1 a),,-1 [~(t + r)] ----4Jl'

21fT

4rr

2Jl'r ar

r

m -I

forn= -2-'

m -I forn= -2-' (22.53)

The theta functions are not needed in (22.53) because the arguments of the delta functions already meet the restrictions imposed by the theta functions. The two functions in (22.53) have an interesting physical interpretation. Green's functions are propagators (of signals of some sort), and Glret)(r, t) is capable of propagating signals only for positive times. On the other hand, G;arlv) (r, t) can propagate only in the negative time direction. Thus, if initially (t = 0) a signal is produced (by appropriate BCs), both G;,et)(r, t) and G;arlV)(r, t) work to propagate it in their respective time directions. It may seem that G;adv)(r, t) is useless because every signal propagates forward in time. This is true, however, only for classical events. In relativistic quantum field theory antiparticles are interpreted mathematically as moviug in the negative time direction! Thus, we cannot simply ignore G;adV)(r, t). In fact, the correct propagator to choose in this theory is a linear combination of G;arlv) (r, t) and G;ret)(r, t), called the Feynman propagator (see [Wein 95, pp. 274-280]). The foregoing example shows a subtle difference between Green's functions for second-order differential operators in one dimension and in higher dimensions. We saw in Chapter 20 that the former are continuous functions in the interval on which they are defined. Here, we see that higher dimensional Green's functions are not only discontinuous, but that they are not even functions in the ordinary sense; they contain a delta function. Thus, in general, Green's functions in higher dimensions ought to be treated as distributions (generalized functions).

22.5

The Eigenfunction Expansion Technique

Suppose that the differential operator Lx,defined in a domain D with boundary aD, has discrete eigenvalues P.n}~l with corresponding orthonormal eigenfunctions {um(x)}~~l' These two sets may not be in one-to-one correspondence. Assume that the U m (xl's satisfy the same BCs as the Green's function to be defined below. Now consider the operator Lx - A1, where A is different from all An'S. Then, as in the one-dimensional case, this operator is invertible, and we can define its Green's function by (Lx - A)G,(X, y) = ~(x - y) where the weight function is set equal to one. The completeness of {un(x)}~=l implies that 00

~ (x - y) = L Un (x)U~ (y) n=l

00

and

G,(x, y) = Lan(y)Un(x). n=l

22.5 THE EIGENFUNCTION EXPANSION TECHNIQUE

637

Substituting lbese two expansions in the differential equation for GF yields 00

00

LO.n - A)an(y)un(x) = LUn(X)U:(y). n=l

n=l

The ortbononnality oflbe Un's gives an(y) = u:(Y)!(An - A). Therefore, _ ~ un(x)u:;(y) ) G ).x'Y-L.-J ( . n=l An - A

(22.54)

In particular, if zero is not an eigenvalue of Lx, its Green's function can be written

as ) _ ~ un(x)u:;(y) X,Y-t=1 An .

G(

(22.55)

This is an expansion of lbe Green's function in terms of lbe eigenfunctions of Lx. I! is instructive to consider a formal interpretation of Equation (22.55). Recall lbat lbe spectral decomposition lbeorem perntits us to write f(A) = Li f(Ai)Pi for an operator A wilb (distinct) eigenvalues Ai and projection operators Pi. Allowingrepetitionofeigenvaluesinlbe sum, we may write f(A) = Ln f(A n) Iun) (unl, where n counts all lbe eigenfunctions corresponding to eigenvalues. Now, let f(A) = A- t . Then

or, in "matrix element" form,

This last expression coincides wilb lbe RHS of Equation (22.55). Equations (22.54) and (22.55) demand lbat lbe un(x) form a complete discrete ortbonormal set. We encountered many examples of such eigenfunctions in discussing Sturm-Liouville systems in Chapter 19. All lbe S-L systems lbere

were, of course, one-dimensional. Here we are generalizing the S-L systemto m dimensions. This is not a lintitation, however, because-for lbe PDEs of interestlbe separation of variables reduces an m-dimensional POE to m one-dimensional ODEs. If lbe BCs are appropriate, lbe m ODEs will all be S-L systems. A review of Chapter 19 will reveal lbat homogeneous BCs always lead to S-L systems. In fact, Theorem 19.1.1 guarantees lbis claim. Since Green's functions must also satisfy homogeneous BCs, expansions such as lbose of (22.54) and (22.55) become possible.

638

22. MULTIOIMENSIONAL GREEN'S FUNCTIONS: APPLICATIONS

22.5.1. Example. As a concrete example, let us obtain an eigenfunction expansion of the GF of the two-dimensional Laplacian, V2 = a2jax 2 + a2ja y 2 , inside the rectangular region 0 ::: x:::: a,a:::: v s bwithDirichletBCs.SincetheGFvanishesattheboundary,the eigenvalue problembecomes V 2 u = AU with u(O, y) = uta, y) = u(x, 0) = u(x, b) = O. The method of separation of variables gives the orthonormal eigenfunctions"

~ sin (n; x) sin (m 7r y)

Umn(X, y) =

b

whose corresponding eigenvalues are

A

mn = _

for m,n

1,2, ... ,

=

[(n;)2 + (mb )2]. 1r

Inserting theeigenfunctions andthe eigenvalues in Equation (22.55), we obtain G( ') _ G( . ' ') _ ~ umn(x, y)umn(x', y') r, r x, y, x ,y - m~l A mn

4

00

sin

= - ab m,n=l L

n1r ) sin (mrr). bY sm (nrr -;;x ,) sin (mrr bY,) (-;;x (nrr)2 (mrr)2 + a

b

wherewe changedx to r and y to r', Note thatthe eigenvaluesareneverzero;thus,G(r, r") is well-defined. III

In the preceding example, zero was not an eigenvalue of Lx. This condition must hold when a Green's function is expanded in terms of eigenfunctions. In physical applications, certain conditions (which have nothing to do with the Bes) exclude the zero eigenvalue automatically when they are applied to the Green's function. For instance, the condition that the Green's function remain finite at the origin is severe enough to exclude the zero eigenvalue.

22.5.2. Example. Let us consider the two-dimensional Dirichlet BVP V 2 u = t, with u = 0 on a circle of radius a. If we consider only the BCs and ask whether zero is an

eigenvalue of V2, the answer will be yes, as the following argument shows. The most general solution to the zero-eigenvalue equation, V 2u = O. in polarcoordinatescanbe obtained by the method of separation of variables: 00

u(p, rp) = A

+ B lup +

00

L (bnP"

+ b~p-n) cosnrp +

n~l

L(cnP" n~l

+ c~p-n) sinnrp. (22.56)

Invoking the Be gives 0= u(a, cp)

00

00

n=l

n=l

= A + BIna + L(bnan +b~a-n) cosncp + L(cnan + c~a-n)sinncp,

which holds for arbitrary rp if and only if A

= -Blua,

b'n = - bn 2n Q

,

4Theinner product is defined as a doubleintegral overtherectangle.

22.5 THE EIGENFUNCTION EXPANSION TECHNIQUE

639

Substituting in (22.56) gives u(p, q;)

= Bin

m+ f

n=l

2 (pn - a : ) (h n cosnq; + en sinnq;).

(22.57)

p

Thus, if we demand nothing beyond the Bes, V2 will have a nontrivial eigen-solution corresponding to the zeroeigenvalue, givenby Equation (22.57). Physical reality, however, demands that u(p, q;) be well-behaved at the origin. This condition sets B, b~. and c~ of Equation (22.56) equal to zero. The Bes then make the remaining coefficients in (22.56) vanish. Thus, the demand that u(p, q;) be well-behaved at p = 0 turns the situation completely around and ensures the nonexistence of a zero eigenvalue for theLaplacian, which in tumguarantees the existenceof a GF. III

In many cases the operator Lx as a whole is not amenable to a full StnrmLiouville treatment, and as such will not yield orthonormal eigenvectors in terms of which the GF can be expanded. However, it may happen that Lx can be broken up into two pieces one of which is an SoL operator. In such a case, the GF can be found as follows: Suppose that Ll and L2 are two commuting operators with L2 an SoL operator whose eigenvalues and eigenfunctions are known. Since L2 commutes with LI, it can be regarded as a constant as far as operations with (and on) Ll are concerned. In particular, (Ll + L2)G = 1 can be regarded as an operatnr equation in LI alone with L2 treated as a constant. Let XI denote the subset of the variables on which L1 acts, and let X2 denote the remainder of the coordinates. Then we can write8(x-y) = 8(Xl -n)8(X2 -Y2). Now let Gl(Xl, n; k) denote the Green's function for LI + k, where k is a constant. Then it is easily verified that (22.58)

In fact, (LI

+ L2)G(X, y) = .[(Ll + L2)Gl(Xl, YI; L2)]8(X2 -

Y2).

=O(Xt-YI) by definition of Gl

Once GI is found as a function of L2, it can operate on 8(X2 - Y2) to yield the desired Green's function. The following example illustrates the technique.

22.5.3. Example. Let us evaluate the Dirichlet GF for the two-dimensional Helmholtz operator V 2 - k 1 in the infinite strip 0 ::::: x ::::: a, -(Xl < Y < 00. Let Ll = a2/a y2 - k 2 and L2 = a2 lax2 • Then, G(r, r ') '" G(x, x', y, y') = Gl (Y, y'; L2)~(X - x'}, where (d 2 Idyl ~ /-,,2jGl = ~(y - y'), 1'2 '" k 2 - L2, and Gl (y = -00) = Gl (y = 00) = O. The GF Gl can be readily found (see Problem 20.12): e-~IY-Y'I

2/k2 ~ L2

.

640

22. MULTIDIMENSIONAL GREEN'S FUNCTIONS: APPLICATIONS

The full GFis then , e-Jk

G(r,r)=

2- L2IY-Y'I)

~

(

2yk2 - L2

r

~(x-x).

(22.59)

The operator Lzconstitutes an S-L system with eigenvalues An = -(nrrx/a)2 andeigenfunctions Un (x) = ,fITa sin(m,. x/a) where n = I, 2, .... Therefore, the delta function 8(x - x') canbe expanded in terms ofthese eigenfunctions:

(mr ) . (mr ')

'( , 2 ~ . ax-x)=-L-sm-xsm-x. a n=l

a

a

As JL = Jk 2 - Lz actson the delta function, Lz operates on the first factor in the above expaosion aodgives An. Thns, L2 in Equation (22.59) cao he replaced by -(nnx/a)2, aod we have

, I ~( G(r,r)=-- La n=1

e-Jk2+(m,x/aj2IY-Y'I) . (nn ) . (nn ') - sm - x sm - x . 2 + (nnx/a)2 . a a

Jk

II

Sometimes it is convenient to break an operator into more than two parts. In fact, in some cases it may be advantageous to define a set of commuting selfadjoint (differential) operators {Mj} such that the full operator L can be written as L = Lj LjMj where the differential operators {Lj} act on variables on which the Mj have no action. Since the Mj 's commute among themselves, one can find simultaneous eigenfunctions for all of them. Then one expands part of the delta function in terms of these eigenfunctions in the hope that the ensuing problem becomes more manageable. The best way to appreciate this approach is via an example. 22.5.4. Example. Letus consider theLaplacian in spherical coordinates, 'V2u =la(2aU) -- r -

r2 ar

ar

2u] a
I [a(. au) +a- . +--- sm82

r sin e ae

ae

If we introduce MtU

= U,

Lt U = Lzu =

~~ (r 2aU) , 2 r aT

ar

I

"2u, r

(22.60)

the Laplacian becomes V2 = L1Ml + LzMz- The mutual eigenfunctions of Ml and Mz are simply thoseof Ml' whichis (thenegative of) the angular momentum operator discussed in Chapter 12,whoseeigenfunctions arethespherical harmonics, Wethus haveM2 YZ m (8, cp) = -1(1 +

l)Ylm(e,
Letus expand the Green's function in terms of the spherical harmonics:

G(r, r') =

L 8Im(r; r', e',
641

22.6 PROBLEMS

Wealso writethedeltafunction as 8( _ ') r

r

= 8(r -

r')8(0 - O')8(rp - rp') . t2 . ()' r

= 8(r

sm

- r') r

t2

"y: LI,m

1m

(0

)Y' (0' ') ,cp 1m ,cp,

where we haveused the completeness of the spherical harmonics. Substituting all of the above in '1 20 (r , r') = 8(r - r'), we obtain ' '1 2 0 (r , r') = (L)M)

+ L2M2) L

gIm(r; r', 0', rp')Ylm(O, rp)

I,m = L{[(LI -1(1

+ I)L2]gIm(r; r', 0', rp'»)Ylm(O, rp)

t.m • , , = 8(r -12r'»" L. Ylm(O ,rp )Ylm(O, rp).

r

t.m

The orthogonality of the Ylm(0, rp) yields [(LI -1(1

+ I)L2]glm(r; r , ,0," rp) =

8(r - r') * , , 2 Ylm(O, rp).

r'

This shows thai the angular part of glm is simply YI';,,(8', rp'). Separaling this from the dependence on r andr' andsubstituting for Ll andL2, we obtain

~~ (r2dglm) _ 1(1 + I) glm = 8(r - r'), ,2 dr dr,2 ,2

(22.61)

where this last gim is a function of r and r' only..The techniques of Chapter 20 can be employed 10 solve Equation (22.61) (see Problem 22.29). III

The separation of the full operator into two "smaller" operators can also be used for cases in which both operators have eigenvalues and eigenvectors. The result of such an approach will, of course, be equivalent to the eigeufunction-expansion approach. However, there will be an arbitrariness in the operator approach; Which operator are we to choose as our LI? While in Example 22.5.3 the choice was clear (the operator that had no eigenfunctions), here either operator can be chosen as LI. The eusuing GFs will be 'equivalent, and the series representing them will be convergent, of course. However, the rate of couvergence may be different for the two. It turns out, for example, that if we are interested in G(x, y; x', y') for the two-dimensioual Laplacian at points (x, y) whose y-coordiuates are far from s'. then the appropriate expansion is obtained by letting LI = a2 ja y 2 , that is, an expansion in terms of x eigenfunctions. On the other hand, ifthe Green's frmctiou is to be calculated for a poiut (x, y) whose x-coordinate is far away from the singular point (x', y'), then the appropriate expansion is obtained by letting L) = a2 jax 2 •

22.6 Problems 22.1. Find the GF for the Ditichlet BVP iu two dimensious if D is the UHP and aD is the x-axis.

642

22. MULTIDIMENSIONAL GREEN'S FUNCTIONS: APPLICATIONS

22.2. Add f(r") to H(r, r") in Example 22.1.2 and retrace the argument given there to show that f(r") = O. 22.3. Use the methnd of images to find the GF for the Laplacian in the exterior region of a "sphere" of radius a in two andthree dimensions. 22.4. Derive Equation (22.7) from Equation (22.6). 22.5. Using Equation (22.7) with p = 0, show that if g(e', rp') = Vo, the potential at any point inside the sphere is Vo. 22.6. Find the BC that the GF must satisfy in order for the solution u to be representable in terms of the GF when the BC on u is mixed, as in Equation (22.10). Assume a self-adjoint SOLPDO of the elliptic type, and consider the two cases a(x) 01 0 and f3(x) 01 0 for x E aD. Hint: In each case, divide the mixed BC equation by the nonzero coefficient, substitute the result in the Green's identity, and set the coefficient of the u term in the D integral equal to zero.

a

22.7. Show that the diffusion operator satisfies Lx"G(x, y; t -

'J =

8(x - y)8(t - c).

Hint: Use

ae

- ( t - r) = 8(t - c).

at

22.8. Show that for m = 3 the expression for G,(r) given by Equation (22.36) reduces to G,(r) = -e-p.r 1(4rr:r). 22.9. The time-independent Schrodinger equation can be rewritten as

where k 2 = 2/1o£/fj,2 and /10 is the mass of the particle. (a) Use techniques of Section 21.4 to write an integral equation for 1jI. (b) Show that the Neumann series solution of the integral equation converges ouly if

(c) Assume that the potential is of Yukawa type: V(r) = g2e- " [r . Find a condition between the (bound state) energy and the potential strength g that ensures

convergence of theNeumann series. 22.10. Derive Equation (22.29).

22.6 PROBLEMS

643

22.11. Derive Equation (22.31) using the procedure outlined for parabolic equations. 22.12. (a) Show that the GF for the Hebrtholtz operator V 2 +/L2 in two dimensions is

where H(r, r') satisfies the homogeneous Helmholtz equation. (b) Separate the variables and use the fact that H is regular at r = r' to show that H can be written as 00

H(r, r') =

L I n(w)[an (r') cos nO + bn(r') sin nO]. n=O

(c) Now assume a circular boundary of radius a and the BC G(a, r') = O.in which a is a vector from the origin to the circular boundary. Using this BC, show that ao(r') = an(r') = b n (r') =

i (1Jr n(t) (/L"/a2 + r 12 - 2ar' cos(O - 0')) dO, 8JfJO(/La) i o 0

i

(1Jr

4Jf In(/La) io

HJt) (/L,,/a 2 + r12 _ 2ar'cos(0 - 0')) cos nO dO,

t" HJt) (/L"/a 2 + r,2 -

i 4Jf In(/La) io

2ar' cos(O - 0')) sin nO dt).

These equations completely determine H(r, r') and therefore G(r, r'),

22.13. Use the Fourier transform technique to find the singular part of the GF for the diffusion equation in one and three dimensions. Compare your results with that obtained in Section 22.4.3. 22.14. Show directly that both G;ret) and G;adv) satisfy V2 G = 8(r)8(/) in three dimensions. 22.15. Consider a rectangular box with sides a, b, and c located in the first octant with one corner at the origin. Let D denote the inside of this box. (a) Show that zero cannot be an eigenvalue of the Laplacian operator with the Dirichlet BCs on aD. (b) Find the GF for this Dirichlet BVP. 22.16. Find the GF for the Helmholtz equation (V 2 o .s x ::: a, 0 s Y :5 b.

+ k2 )u =

0 on the rectangle

22.17. Find the singular part of the one-dimensional GF for the operator ad 2jdx2 + b, where a > 0 and b < O.

644

22. MULTIDIMENSIONAL GREEN'S FUNCTIONS: APPLICATIONS

22.18. Calculate the GF of the two-dimensional Laplacian operator appropriate for Neumann Bes on the rectangle 0 :5 x :5 a, 0 :5 Y :5 b. 22.19. Find the three-dimensional Dirichlet GF for the Helmholtz operator \72 _k2 in the half-space z "': O. 22.20. Find the three-dimensional Neumann GF for the Helmholtz operator \72 k 2 in the half-space z :5 O. 22.21. Using the integral form of the Schriidinger equation in three dimensions, show that an attractive delta potential V (r) = - Vo8 (r - a) does not have a bound state (E < 0). Contrast this with the result of Example 21.4.1. 22.22. By taking the Fourier transform of both sides of the integral form of the Schriidinger equation, show that for bound-state problems (E < 0), the equation in "momentum space" can be writtenas 'iJ(p) = where

/(2

(27<~~21i2 C2 ~ p2)

J

V(p - q)'iJ(q)d

3q,

= -2JLElIi2.

22.23. Write the bound-state Schrodinger integral equation for a non-local potential, noting that G(r, r') = e-
M: { 27<"2 JlR3

e, ( d 3r,e-
[r - r'l

(22.62)

(b) Taking U(r) = e-a'lr, show that the condition in (22.62) becomes

47< JLg2 [ I ] _ 1 ali2 (a + /()2 - . (c) Since /( > 0, prove that the equation in (b) has a unique solution only if g2 > 1i2a21(47
2 2) E = _1i [(47
_

a]2

22.24. Repeat calculations in Sections 22.4.1 and 22.4.2 for m = 2. 22.25. In this problem, the dimension m is three. (a) Derive the following identities:

22.6 PROBLEMS

645

2 V28(1 ± r) = 8"(1 ± r) ± -8'(1 ± r), dl r where E(I) = 0(1) - O(-t)o (b) Use the results of (a) to show that the GF [Equation (22.47)] derived from the principal value of the ill integration for the wave equation in three dimensions satisfies only the homogeneous PDE. Hint: Use V 2(I/r) = 411"8(r). dE(I) = 28(1),

22.26. Calcnlate the retarded GF for the wave operator in two dimensions and show that it is equal to G~ret) (r, I) = 0(1)/(211"./1 2 - r 2). Now use this resnlt to obtain the GF for any even number of dimensions:

] G(ret)(r I) _ 11(1) ( __I_~)n-j [. I , , - 211" 211"r ar ./12 _ r2

for n = m/2.

22.27. (a) Find the singniar part of the retarded GF and the advanced GF for the wave equation in three dimensions using Equations (22.48) and (22.49). Hint: iJ/2(kr) = ./2/nkrsinkr. (b) Use (a) and Equation (22.51) to show that

lim { .....0

[r 2

I

+ (-il + E)2]

_

[r 2

I

+ (it + E)2]

} = _ i1l" [8(1 + r) _ 8(1 _ r)] . r

22.28. Show that the eigenfunction expansion of the GF for the Dirichlet BVP for the Laplacian operator in two dimensions for which the region of interest is the interior of a circleof radius a is

where EO = ~ and En = I for n 2: I, and use has been made of Problem 14.39. 22.29. (a) Complete the calculations of Example 22.5.4, and (b) find the GF for the Laplacian with Dirichlet BCs on two concentric spheres of radii a and b, with a < b. (c) Consider the case where a -> 0 and b -> 00 and compare the resnlt with the singular part of the GF for the Laplacian. 22.30. Solve the Dirichlet BVP for the operator V2 - k 2 in the region 0 :0: x :0: a, o :0: y :0: b, -00 < z < 00. Hint: Separate the operator into Lj and L2. 22.31. Solve the problem of Example 22.5.1 using the separation of operator tecimique and show that the two resnlts are equivalent. 22.32. Use the operator separation technique to calcnlate the Dirichlet GF for the two-dimensional operator V2 - k 2 on the rectangle 0 :0: x :0: a, 0 :0: y :0: b. Also obtain an eigenfunction expansion for this GF.

646

22. MULTIDIMENSIONAL GREEN'S FUNCTIONS; APPLICATIONS

22.33. Use the operator separation techniqne to find the three-dimensional Dirichlet GF for the Laplacian in a circular cylinder of radius a and height h. 22.34. Calculate the singular part of the GF for the three-dimensional free Schrodinger operator

a ft2 ift- - _';72 2fJ-

at

22.35. Use the operator separation technique to show that (a) the GF for the Helmholtz operator ';72 + k 2 in three dimensions is I

00

G(r, r') = -ik

L L

jl(kr<)hl (kr»Ylm «(J, CP)Y/;n«(J', cp'),

l=Om=-1

where r -c (r» is the smaller (larger) of r and r' and jl and hi are the spherical Bessel and Hankel functions, respectively. No explicit BCs are assumed except that there is regularity at r = 0 and that G(r, r') --> 0 for [r] --> 00. (h) Obtain the identity iklr-r'l

e

4][lr - r'l

I

00

= ik

L L I~O m~-I

jl(kr<)hl(kr»Ylm«(J, CP)Y/;II«(J', cp').

(c) Derive the plane wave expansion [see Equation (19.26)]

ei k .r = 4][

00

I

L L

i l jl(kr)YI~,«(J', cp')Ylm«(J, cp),

[=0 m=-l

where s' andcp' are assumed to be the angular coordinates of k. Hint: Let [r'] --> 00, and use

[r - r'] = (r'2

+ r2 -

r' r

2r· r')1/2 --> r' - - ' r'

and the asymptotic formula h?) (z) ---> (ljz)e i [z+(I+ I )(1r/2)], valid for large z.

Additional Reading 1. Economou, E. Green's Functions in Quantum Physics, Springer-Verlag, 1983. Emphasizes applications of Green's function in quantum mechanics, especially solid-state physics. 2. Jackson, J. Classical Electrodynamics, 2nd ed., Wiley, 1975. Many examples of Green's function techniques used in electromagnetism, in particular the advanced and retarded Green's functions as used in electromagnetic radiation theory.

22.6 PROBLEMS

647

3. Roach, G. Green's Functions, Van Nostrand, 1970. Treats eigenfunction expansion of Green's functions and gives many examples.

Part VII

Groups and Manifolds

_

23

_

Group Theory

The tale of mathematics and physics has been one of love and hate, of harmony and discord, and of ftiendship and animosity. From their simnltaneous inception in the shape of calculus in the seventeenth century, through an intense and interactive development in the eighteenth and most of the nineteenth century, to an estrangement in the latter part of the nineteenth and the beginning of the twentieth century, mathematics and physics have experienced the best of times and the worstof times. Sometimes,as in the case of calculus,nature dictates a mathematical dialect in which the narrative of physics is to be spoken. Other times, man, building upon that dialect, develops a sophisticated language in which-c-as in the case of Lagrangian and Hamiltonian interpretation of dynamics-the narrative of physics is set in the most beautiful poetty. But the happiest courtship, and the most exhilarating relationship, takes place when a discovery in physics leads to a development in mathematics that in tum feeds back into a better understanding of physics, leading to new ideas or a new interpretation of existing ideas. Such a state of affairs began in the 1930s with the advent of quantum mechanics, and, after a lull of about 30 years, revived in the late 1960s. We are fortunate to be wituesses to one of the most productive collaborations between the physics and mathematics communities in the history of both. I! is not an exaggeration to say that the single most important catalyst that has facilitated this collaboration is the idea of symmetry the study of which is the main topic of the theory of groups, the subject of this chapter. Although group theory, in one form or another, was known to mathematicians as early as the beginning of the nineteenth century, it found its way into physics ouly after the invention of quantum theory, and in particular, Dirac's interpretation of it in the language of

652

23. GROUP THEORY

transformation theory. Eugene Wigner, in his seminal paper! of 1939 in which he applied group theoretical ideas to Lorentz transformations, paved the way for the marriage of group theory and quantum mechanics. Today, in every application of quantum theory, be it to atoms, molecules, solids, or elementary particles such as quarks and leptons, group-theoretical techuiques are indispensable.

23.1 Groups The prototype of a group is a transformation group, the set of invertible mappings of a set onto itself. Let us elaborate on this. First, we take mappings because they are the most general operations performed between sets. From a physical standpoint, mappings are essential in understanding the symmetries and other basic properties of a theory. For instance, rotations and translations are mappings of space. Second, the mappings oughtto be on a single set, because we want to be able to compose any given two mappings. We cannot compose f : A ---> B and g : A ---> B, because, by necessity, the domain of the second must be a subset of the image of the first. With three sets, and A J.. B, B .!, C, even if the composition fog is defined, g 0 f will not be. Third, we want to be able to undo the mapping. Physically, this means that we should be able to retrace our path to our original position in the set. This can happen only if all mappings of interest have an inverse. Finally, we note that composing a mapping with its inverse yields identity. Therefore, the identity map must also be included in the set of mappings. We shall come back to transformation groups frequently. In fact, almost all groups considered in this book are transformation groups. However, as in our study of vector spaces in Chapter I, it is convenient to give a general description of (abstract) groups. . Group defined

23.1.1. Definition. A group is a set G together with an associative binary operation G x G ---> G called multiplication---
1. There exists a unique eiemenr e E G called the Ulentiry such that e * g = g*e=g. 2. For every element g E G, there exists an element g-l, called the inverse of g, such that g * g-I = g-I * g = e. To emphasize the binary operation ofa group, we designate it as (G, *). order of a group

If the underlying set G has a finite number of elements, the group is called finite, and its number of elements, denoted by 1G I, is called the order of G. We can also have an infinite group whose cardinality can be countable or continuous.

~E. P. Wigner, "On the UnitaTf Representations of the Inhomogeneous Lorentz Group,"Ann. afMath. 40 (1939) 149-204. To distinguish between identities of different groups, we sometimes write ec for the identity of the group G.

23.1 GROUPS

653

Given an element a E G, we write

ak =a*a*···*a, '-,-'

a-m es a-I *0- 1 * ... *a- 1. m times

k times

and note that

for all i, j

E

Z.

Evariste Galois (1811-1832) was definitely not the stereotypically dull mathematician, quietly creating theorems and teaching students. He was a political firebrand whose life ended in a mysterious duel when he was only 21 years old. An ardent republican, he was in the unfortunate position of having Cauchy, an ardent royalist, as the only French

mathematician capable of understanding the significance ofhis work. His professional accomplishments (fewer than 100 pages, much of which was published posthumously) received the attention they deserved many years later. It is truly sad to realize that for decades, work from the man credited with the foundation of group theory were lost to the world of mathematics. Galois's early years were relatively happy. His father, a liberal thinker known for his wit, was director of a boarding school and later mayor of Bourg-la-Reine, Galois's mother took charge of his early education. A stubborn, eccentric woman, she mixed classical culture with a fairly stem religious upbringing..The young Galois entered the College Louis-Ie-Grand in 1823, but found the harsh discipline imposed by church and political authorities difficult to bear. His interest in mathematics was sparked in class by Vernier, but Galois quickly tired of the elementary character of the material, preferring instead to read the more advanced original works on his own. After a flawed attempt to solve the-general fifth-order equation, Galois submitted a paper to the Academie desSciences in which he described the definitive solution with the aid of group theory, of which the young Galois can be considered the creator. However, this strong initial foray into the frontiers of mathematics was accompanied by tragedy and setback. A few weeks after the paper's submission, his father committed suicide, which Galois felt was largely to be blamed on those who politically persecuted his father. A month later the young mathematician failed the entrance examination to the Ecole Polytechnique, largely due to his refusal to answer in the fonn demanded by the examiner. Galois did gain entrance to a less prestigious school for the preparation of secondary-school teachers. While there he read some of Abel's results (published after Abel's death) andfound that they contained some of the results he had submitted to the Academy including the proof of the impossibility of solving quintics. Cauchy, assigned as the judge for Galois's paper, suggested that he revise it in light of this new information. Galois instead wrote an entirely new manuscript and submitted it in competition for the grand prix in mathematics. Tragically, the manuscript was lost on the death of Fourier, who had been assigned to examine it, leaving Galois out of the competition. These events, fueled by a later, unfair dismissal of another of his papers by Poisson, seem to have driven Galois toward political radicalism and rebellion during the renewed turmoil then plaguing France. He was arrested several times

654

23. GROUP THEORY

for his agitations, althoughhe continued work on mathematicswhile in custody. On May 30, 1832, he was wounded in a duel with an unknown adversary, the duel perhaps caused

by an unhappylove affair. Hisfuneral three days later sparkedriots that raged throughParis in the days that followed. The delay in recognition of the true scope of Galois's scant but amazing work stemmed partly from the originality ofhis ideas and the lack of competent local reviewers. Cauchy left

Franceafter seeingonly the early parts of Galois's work, and muchof the rest remainedunnoticed until Liouville prepared the later manuscripts for publication a decade after Galois's death. Their true value wasn't appreciated for another two decades. The young mathematician himself added to the difficnlty by deliberately making his wriring so terse that the "established scientists" for whom he had so much disdain could not understand it. Those fortunate enough to appreciate Galois's work found fertile ground in mathematical research, in such fundamental fields as group theory and modem algebra, for decades to come.

23.1.2. Example. The followingareexamplesoffarniliar setsthat havegroupproperties. (a) The set Z of integers under the binary operation of addition forms a group whose identity element is 0 and the inverse of n is -rtt, This group is countably infinite.

(b) The set I-I, +Ij, onder the binary operation of mnltiplication, forms a gronp wbose identity element is 1 and the inverse of each element is itself. This group is finite.

(c) The set (-1, +1, -i, +ij, under the binary operation ofmnltiplication, forms a finIte group whose identity element is I.

(d) The set JR, onder the binary nperationof addition,forms a gronp whose identityelement is 0 and the inverse of r is -r. This group is uncountably infinite.

(e) The setJR+ (1Ql+) of positive real (rational) nnmbers, onder the binary operation of multiplication, forms a group whose identity element is 1 and the inverse of r is 1/ r, This group is uncountably (countably) inlinite. (f) The set C, under the binary operation of addition, forms a group whose identity element is 0 and the inverse of z is -z. This group is uncountably infinite. (g) The uncountablyinfiniteset
is I/z. (h) The uncountably infinite set V of vectors in a vector space, under the binary operation of addition, forms a group whose identity element is the zero vector and the inverse of la)

is -tal. (i) The set of invertible n x n matrices, under the binary operation of multiplication, forms a group whose identity element is the n x n unit matrix. and the inverse of A is A-1. This group is uncountably infinite. II The reader is urged to verify that each set given above is indeed a group.

In general, the elements of a group do not commute. Those groups whose elements do commute are so important that we give them a special name: abelian groups 23.1.3. Definition. A group (G, *) is calledabelian or commutative if aeb = bea defined for all a, bEG. It is common to denote the binary operation ofan abelian group

by+. All groups of Example 23.1.2 are abelian except the last.

23.1 GROUPS

655

23.1.4. Example. Let A be a vector potential that gives rise to a magnetic fietd B. The set of transformations of A that give rise to the same B is an abelian group. In fact, such transformations simply add the gradient of a function to A. The reader can check the III

details.

symmetric or permutation group

The reader may also verify that the set of invertible mappings f : S -+ S, i.e., the set of transformations of S, is indeed a (nonabelian) group. If S has n elements, this group is denoted by Sn and is called the symmetric gronp of S. Sn is a nonabelian (unless n ::; 2) finite group that has n! elements. An element g of S« is usually denoted by two rows, the top row being S itself-usually taken to be I, 2, ... , n-and the boltom row its image under g. For exannple, g sa ~ ~ is an element of S4 such that g(l) = 2, g(2) = 3, g(3) = 4, and g(4) = I. Consider two groups, the set of vectors in a plane y), +) and the set of complex numbers (C, -t-), both under addition. Although these are two different groups, the difference is superficial. We have seen similar differences in disguise in the context of vector spaces and the notion of isomorphism. The sanne notion applies to group theory:

(i 1)

«x,

homomorphism, isomorphism, and automorphism

23.1.5. Definition. Let (G, *) and (H, 0) be groups. A homomorphism f : G-+ H is a map such that

f(a *b) = f(a)

0

feb)

Va,b

E

G.

An isomorphism is a homomorphism that is also a bijection. Two groups are isomorphic, denoted by G ~ H, if there is an isomorphism f : G -+ H. An isomorphism ofa group onto itself is called an automorphism. An inmnediate consequence ofthis definition is thatf(eG) = en and f(g-l) = [f(g)]-l (see Problem 23.9). trivial homomorphism

23.1.6. Example. Let G be any group and (I} the multiplicativegroup cousistiug of the singlenumber 1. It isstraightforward to showthat 1 : G --+ (I}, givenby (the ouly function available!) l(g) = I for all g EGis a homommphism. This homomorphism is called the trivial (or sometimes, symmetric) homomorphism. II

«x,

general linear group

The establishment of isomorphism f : 1R2 -+ C between y), -l-), and (C, +) is trivial: Just write f(x, y) = x + iy. A less trivial isomorphism is the exponential map, exp : (JR, +) -+ (JR+,.). The reader may verify that this is a homomorphism (in particular, it maps addition to multiplication) and that it is one-to-one. We have noted that the set of invertible maps of a set forms a group. A very important special case of this is when the set is a vector space V and the maps are all linear

23.1.7. Box. The general linear group of a vector space V, denoted by

GL(V), is the set ofall invertible endomorphisrns ofV. In particular, when

V = C", we usually write GL(n, iC) insteadofGL(iC") with similar notation forlR.

656

23. GROUP THEORY

group multiplication table

It is sometimes convenientto display a finite group G = {g;})~\ as a IGI x IGI table, called the grnup multiplication table, in which the intersection of the ith row and jth column is occupied by gi * gj. Because of its trivial multiplication, the identity is usually omitted from the table.

23.2 Subgroups It is customary to write ab instead of a * b. We shall adhere to this convention, but restore the * as necessary to avoid any possible confusion. subgroup defined

23.2.1. Definition. A subset S ofa group G is a subgroup of G if it is a group in its own right under the binary operation of G, i.e., if it contains the inverse ofall its elements as well as the product ofany pair of its elements. It follows from this definition that e E S. It is also easy to show that the intersection of two subgroups is a subgroup (Problem 23.2).

trivial subgroup

23.2.2. Example. ExAMPLES OF SUBGROUPS I. For any G, the subset {e}, consistingof the identity alone, is a subgroupofG called the trivial subgroup of G. 2. (:I:, -t-) is a subgroupof (R, +). 3. The set of even integers (but not odd integers) is a subgroup of (:I:, +). In fact, the set of all multiples of a positive integer m, denoted by Zm, is a subgroup of Z. It turns out that all subgroupsof Z are of this form.

special linear group

4. The subset of GL(n, C) consisting of transformations that have unit determinant is a subgroup of GL(n, C) because the inverse of a transformation with unit determinant also has unit determinant,and the product of two transformations with unit determinants has unit determinant.

23.2.3. Box. The subgroup of GL(n, C) consisting of elements having unit determinant is denoted try SL(n, C) and is called the special linear group.

unitary, orthogonal, special unitary, and speciai orthogonal groups

5. The set of uoitary transformations of C", denoted by U(n), is a subgroup of GL(n, C) because the inverse of a unitary transformation is also unitary and the product of two unitary transformations is unitary.

23.2.4. Box. The set of unitary transformations U (n) is a subgroup of GL(n, iC) and is called the unitary group. Similarly, the set of orthogonal transformations of~n is a subgroupofGL(n, ~). It is denoted by O(n) and called the orthogonal group.

23.2 SUBGROUPS

657

Each of these groupshas a specialsubgroupwhose elementshave unit determinants. These are denoted by SU(n) aod SO(n), aod called special unitary group aod special orthogonalgroup, respectively. The latter is also called the group of rigid rotations of R" . 6. Let x, Y E JR n , and define an inner product on R" by

x- Y = Denote

-XIYl - ... - xpYp

+ Xp+lYp+l + ... +xnYn·

the subset of G L(n, JR) that leaves this inner product invariant by3 0 (p, n -

pl. Then O(p, n - p) is a subgroupof GL(n, R). The set oflinear transformations among O(p, n - p) that have determinaot t is denoted by SO(p, n - pl. The Lorentz and Poincare groups symplectic group

specialcase of p = 0 givesus the orthogonaland specialorthogonalgroups." When n = 4 and p = 3, we get the inner product of the special theory of relativity, and 0(3, I), the set of Lorentz transformations, is called the Lorentzgroup. If one adds traoslatioos ofli4 to 0(3, 1), one obtalosthe Poincare group, P(3, 1). 7. Let x, y E ]R2n, and J the 2n x 2n matrix (~1 ~), where 1 is the n x n unit matrix. The subset of GL(2n, JR) that leaves xt Jx, called an antisymmetric bilinearform, invariantis a subgroupof GL(2n, Ii) called the symplectic group aod denotedby Sp(2n, R). As we shall see in Chapter 26, the symplectic group is fundamental in the formal treatment of Hamiltonian mechanics.

conjugate Subgroup

8. Let S be a subgroupof G aod g

E

G. Theo it is readily shown that the set

g-lSg ea (g-lsgls E S) is also a subgroup of G. called the subgroupconjugateto Sunder g. or-the subgroup III g.conjugate to S.

When discussing vector spaces, we noted that given any subset of a vector space, one could construct. a subspace out of it by taking all possible linear combinations (natnra! operations of the vector space) of the vectors in the subset. We called the subspace thus obtained the span of the subset. The same procedure is applicable in group theory as well. If S is a subset of a group G, we can generate a subgroup out of S by collecting all possible products and inverses (natural operations of the group) of the elements of S. The reader may verify that the result is indeed a subgroup of G. subgroup generated bya subset

23.2.5. Definition. Let S be a subset of a group G. The subgroup generated by S, denoted by (S), is the union of S and all inverses and products ofthe elements ofS.

cyclic subgroup

In the special case for which S = {aj, a single element, we use (a) instead of ({aJ) and call it the cyclic subgroup generated by a. It is simply the collection of all integer powers of a.

commutator ofgroup elements

23.2.6. Definition. Let G be a group and a, bEG. The commutator ofa and b,

3The reader is warned that what we have denoted by O(p, n - p) is sometimes denoted by other authors by O(n - p, p) or Otn, p) or O(p, n). 4It is customary to write O(n) and SO(n) for 0(0. n) and SO(O. n)

658

23. GROUP THEORY

denoted by [a, b], is [a, b] es aba-1b- 1 .

commutator subgroup of a group

The subgroup (Ua bEG[a, b]) generated by all commutalors of G is called the commutator subgroup of G. The reader may verify that a group is abeliau if aud only if its commutator subgroup is the trivial subgroup, i.e., consists of only the identity element.

centralizer ofan 23.2.7. Definition. Let x E G. The set of elements of g that commute with x, element in G and the denoted by GG(x), is called the centralizer of x in G. The set Z (G) of elements center of G ofa group G that commute with all elements of G is called the center of G. 23.2.8. Theorem. GG(x) is a subgroup ofG and Z(G) is an abelian subgroup of G. Furthermore, G is abelian ifand only if Z(G) = G. Proof. Proof is immediate from the definitions.

kernel ofa homomorphism

D

23.2.9. Definition. Let G and H be groups and let f : G --> H be a homomorphism. Define the kernel of f by ker f sa {x E G I f(x) = e E H}. The reader may check that ker f is a subgroup of G, aud f (G) is a subgroupof H. These are the aualogues of the same concepts encounteredin vector spaces. In fact, if we treat a vector space as au additivegroup, with the zero vector as identity, then the abovedefinitioncoincides with that of linear mappings aud vector spaces. Carrying the aualogy further, we recall that given two subspaces 11 aud W of a vector space V, we denote by 11 + W all vectors of V that cau be written as the sum of a vector in 11 aud a vector in W. There is a similar concept in group theory that is sometimes very useful. 23.2.10. Definition. Let Sand T be subsets of a group (G, *). Then one defines the product ofthese subsets as

In particular, if T consists ofa single element t, then

SH = {sHls

left and right cosets

E S}.

As usual, we shall drop the * and write ST and St. If S is a subgroup, then St is called a right coseP of Sin G. Similarly, tS is called a left coset of Sin G. In either case, t is said to represent the coset.

5 Some authors switchourright andleft in theirdefinition.

23.2 SUBGROUPS

659

23.2.11. Example. Let G = JR.3 treated as an additiveabeliangroup,andlet S be a plane through the origin. Then t + S is S if t E S (see Problem 23.5); otherwise, it is a plane III parallelto S. Iu fact, t + S is simplythe translationof all points of S by t. 23.2.12. Theorem. Any two right (left) cosets ofa subgroup are either disjoint or identical.

Proof Let S be a subgroup of G and sUfPose that x E Sa n Sb. Then x = Sla = S2b with Sl, S2 E S. Hence, ab:" = sl S2 E S. By Problem 23.6, Sa = Sb. The D left cosets can be treated in the same way.

A more "elegant" proof starts by showing that an equivalence relation can be defined by a

[xl

b

{==}

ab- I E S

and then proving that the equivalence classes of this relation are cosets of S. One interpretation of Theorem 23.2.12 is that a and b belong to the same right coset of S if and only if ab- I E S. A second interpretation is that a coset can be represented by anyone of its elements (why?). All cosets (right or left) of a subgroup S have the same cardinality as S itself. This can readily be established by considering the map : S -+ Sa ( : S -+ as) with (s) sa ((s) as) and showing that is bijective. There are many instances both in physics and in mathematics in which a collection of points of a given set represent a single quantity. For example, it is not simply the set of ratios of integers that comprise the set of rational numbers, but the set of certain collections of such ratios; The rational number ~ represents ~, ~, ~, etc. Similarly, a given magnetic field represents an infiuitude of vector potentials each differing by a gradient from the others, and a physical state in quantum mechanics is an infiuite number of wave functions differing from one another by a phase. With the set of cosets constructed above, it is natural to ask whether they could be given an algebraic structure. The most natural such structure would clearly be that of a group; Given as and bS define their product as abS. Wonld this operation turn the set of (left) cosets into a group? The following argument shows that it will, under an important restriction. It is clear that the identity of such a group would be S itself. It is equally clear that we shonld have (b- 1S)(bS) = S, so that (b-ISb)S = S. It follows from Problem 23.5 that we must have b-ISb c S for all bEG. Now replace b withb- I and note thatbSb- 1 c S as well. Lets be an arbitraryelementofS.Thenbsb- 1 s'farsomes' E S,ands b-Is'b E b-1Sb. It follows that S C b- I Sb for all bEG, and, with the reverse inclusion derived above, that S = b- I Sb, This motivates the following definition.

=

=

=

normal subgroup detined

=

23.2.13. Definition, A subgroup N ofagroup G is called normal if N = g-I Ng (equivalently if N g = g N) for all g E G.

660

23. GROUP THEORY

The preceding argument shows that the set of cosets (no specification is necessary since the right and left cosets coincide) of a normal subgroup forms a group: quolienl group

23.2.14. Theorem. If N is a normal subgroup of G, then the collection of all cosets of N, denoted by GIN, is a group, called the quotient group of G by N. We note that all subgroups of an abelian group are automatically normal. and that the only subgroup conjugate to a normal subgroup N is N itself (see Example 23.2.2).

23.2.15. Example. Let G = JR.3 and let S be a plane tbrough the origin as in Example 23.2.11. Since G is abelian, S is automatically normal,and GIS is the set of planesparallel to S. Let ell be a normalto S. Then it is readily seen that

GIS = {ren + Sir E JR.}. Wehave pickedtheperpendicular distance between a planeand S (with sign included) to represent that plane. The reader may check that the quotientgroup G/ S is isomorphicto R, Identifying S with JR. 2, we can write JR.3 jJR.2 ::: R. The cancellation of exponents is quite accidental here! Let G = Z and S = Zm, the set of multiples of the positive integer m. Since Z is abelian, Zm is normal, and Z/Zm is indeeda group, a typicalelementof whichlooks like k + mZ. By adding (or subtracting) multiples of m to k, and using mj + mZ = mZ (see Problem 23.5), we can assumethat 0 .S k ::; m - I. It follows that ZjZm is a finite group.

Furthermore, (kt +mZ) + (k2 +mZ) = kl +k2 +mZ = k+mZ,

where k is theremainder after enough multiples of m havebeen subtracted from kl + k2. One writes kl + k2 == k (mod m). The coset k + mZ is sometimes denoted by k andthe quotientgroup ZjZm by Zm: Zm =

to, T, 2, ... , m -I}.

Zm is a prototype of the finite cyclic groups. It can be shown that every cyclic group of order m is isomorphic to Zm a generator of which is j (recallthatthe binary-operation is addition for Zm). III first isomorphism lheorem

23.2.16. Theorem. (first isomorphism theorem) Let G and H be groups and f : G ---* H a homomorphism. Then ker f is a normal subgroup of G, and G I ker f is isomorphic to f (G).

Proof We have already seen that ker f is a subgroup of G. To show that it is normal, let g E G and x E ker f. Then f(gxg- 1) = f(g)f(x)f(g-l) = f(g)eHf(g-l) = f(g)f(g-l) = f(gg-l)

= f(eG) = en-

23.2 SUBGROUPS

661

It follows that gxg- 1 E ker f. Therefore, ker f isnonnal. We leaveitto the reader to show that


23.2.18. Definition. Let x E G. A conjugate ofx is an element y of G that can be written as y = gxg- 1 with g E G. The set of all elements ofG conjugate to one another is called a conjugacy class. The ith conjugacy class is denoted by Ki, One can check that "x is conjugate to y" is an equivalence relation whose classes are the conjugacy classes. In particular, two different conjugacy classes are disjoint. One can also show that each element of the center of a group constitutes a class by itself. In particular, the identity in any group is in a class by itself. and each element of an abelian group forms a (different) class. Although a normal subgroup N contains the conjugate of each of its elements, N is not a class. The class containing any given element of N will be ouly a proper subset of N (unless N is trivial). The characteristic feature of a normal subgroup is that it contains the conjugacy classes of all its elements. This is not shared by other subgroups, which, in general, contain only the trivial class of the identity element. 23.2.19. Example. Consider the group SO(3) of rotations in three dimensions. Let us denote a rotation by R~ (f), where e is thedirection of the axisof rotation and () is theangle of rotation. A typical member oflbe conjngacyclass of Re(B) is RRe(B)R- 1, where R is somerotation. Let ~ = Re be thevectorobtained byapplying therotation R on e,andnote tbat RRe(B)R- 1.,

allrotations having the sameangle belong to the same conjugacy class

= RRe(B)R- 1R. = RR e• = R. = .',

where we usedthefact that R@e = e becausea rotation leavesits axis unchanged. Thislast statement, applied to theequation above,also shows that RR@(f))R- 1 is arotation about S'. Problem 23.18 establishestbat the angle of rotation associated with RR e(6)R- 1 is 6. We can summarizethis as RR e(6)R- 1 = Re,(B). It follows tbat all rotations having the same anglebelongtothesameconjugacy class ofthe group ofrotations in three dimensions. II

23.2.1 Direct Products The resolution of a vector space into a direct sum of subspaces was a useful tool in revealing its structure. The same idea can also be helpful in studying groups. Recall that the only vector common to the subspaces of a direct sum is the zero vector. Moreover, any vector of the whole space can be written as the sum of vectors taken 6Compare this theorem with the set-theoretic result obtained in Chapter 0 wherethe map XI bijective if Ex; is theequivalence relation induced by f.

Ex;-+

f(X) was shownto be

662

23. GROUP THEORY

from the subspaces of the direct sum. Considering a vector space as a (abelian) group, with zero as the identity and summation as the group operation, leads to the notion of direct product. internal direct product of groups

23.2.20. Definition. A group G is said to be the direct product of two of its subgroups HI and H2, and we write G = HI X H2, if

1. all elements of HI commute with all elements of H2; 2. the group identity is the only element common to both HI and H2;

3. every g

E

G can be written as g = hlh2 with hI

E

HI and h2

E

H2.

It follows from this definition that h I and h2 are unique, and HI and H2 are normal. This kind of direct product is sometimes called internal, because the "factors" HI and H2 are chosen from inside the group G itself. The external direct product results when we take two unrelatedgroups and make a group out of them: external direct product of groups

23.2.21. Proposition. Let G and H be groups. The Cartesian product G x H, called the external direct product of G and H, can be given a group structure by

(g, h) * is'. h') es (gg', hh'). Furthermore,G ~ Gx{eH},H ~ {eG}xH,GxH ~ HxG,andtowithinthese isomorphisms, G x H is the internal direct product ofG x {eH} and leG} x H. The proof is left for the reader.

Niels Henrik Abel (1802-1829) was the second of seven chil-

dren, son of a Lutheran minister with a smallparish of Norwegian coastal islands. In school he received only average marks at first, butthenhis mathematics teacher was replaced by a man only seven years older than Abel. Abel's alcoholic father died in 1820, leaving almostno financial support for his young prodigy, who became responsible for supporting his mother andfamily. His teacher, Holmboe, recognizing his talent for mathematics, raisedmoney from his colleagues to enable Abel to attend Christiania (modem Oslo) University. Heentered theuniversity in 1821,10years after theuniversity was founded, and soonproved himselfworthy of his teacher's accolades. His secondpaper, forexample, contained thefirst example of a solution to an integral equation. Abelthen received a two-year govemmenttravel grant andjourneyedtoBerlin, where he mettheprominent mathematician Crelle, whosoonlaunched what wastobecometheleading Gennan mathematical journal of the nineteenth century, commonly called Crelle'sJournal. From the start, Abel contributed important papers to Crelle's Journal, including a classic paper onpower series, thescopeof whichclearly reflects his desire forstringency. His most important work, alsopublished in thatjournal, wasalengthytreatment of ellipticfunctions in

23.3 GROUP ACTION

663

which Abelincorporated their inverse functions to showthat they are anatural generalization of the trigonometric functions. In laterresearch in this area, Abel found himself in stiff competition withanother youngmathematician, K. G. J. Jacobi. Abelpublished somepapers on functional equations andintegrals in 1823. hi it he gives the first solutionof anintegral equation. In 1824he proved theimpossibility of solving algebraically the general equation of the fifth degreeandpublished it at his own expense hopingto obtain recognition for his work. Despitehis proven intellectual success,Abel neverachieved material success,noteven a permanent academic position. In December of 1828, while traveling by sled to visit his fiance forChristmas, Abelbecameseriously ill anddieda coupleof months later. Ironically, his death from tuberculosis occurred two days beforeCrellewrote with thehappy news of an appointment for Abel at a scientific institute in Berlin. In Abel's eulogy in his journal, Crellewrote: "He distinguished himselfequallyby the purity aud nobility of his characteraud by a rare modestywhichmadehis personcherished to the samedegreeas was his genius."

23.3 Group Action The transformation groups introduced at the beginning of this chapter can be described in the language of abstract groups. left action, right actlon.len invariance and right invariance

23.3.1. Definition. Let G be a group and M a set. The left action of G on M is a map : G x M --> M such that

1. (e,m) = m for all m EM; 2. (g1g2, m) = (gl, (g2, m». One usually denotes (g, m) by g. m or more simply by gm. The right action is defined similarly. A subset N c M is called left (right) invariant if g . mEN (m'g EN)forallg E G,wheneverm E N. any group is isomorphic to a subgroup ofthe group of transformations of an appropriate set

23.3.2. Example. lfwe define fg : M --> M by fg(m) es <1>(g, m) = g. m, then I g is recognized as a transformation of M. The collectionof such transformations is a subgroup of the set of all transformations of M. Indeed, the identity transformation is simply [e, the inverse of /g is /g-I, andthe(associative) law of composition is fg 1 0 fg 2 = /gI82' There is a general theorem in group theory stating thatany group is isomorphic to a subgroup of IIIlII the group of transformations of an appropriate set.

orbit, stabilizer; transitive action and

23.3.3. Definition. Let G acton M and let mo E M. The orbitofmo, denoted by Gmo, is

effective action

Gmo

= {m E M Im = gmo

for some g E

Gl·

The action is called transitive if Gmo = M. The stabilizer ofmo is Gmo = {g E G I gmo = mol· The group action is called effective if gm = m for all m E M implies that g = e.

664

23. GROUP THEORY

Stabilizer is a subgroup.

The reader may verify that the orbit Gmo is the smallest invariant subset of M containing mo, and that

23.3.4. Box. The stabilizer ofmo is a subgroup of G, which is sometimes called the little group of G at mo.

A transitive actiou is characterized by the fact that given any two points ml, m2 E M, one can find e g e G such that m2 = gm«. 23.3.5. Example. Let M = lit 2 and G = 80(2), the ptanarrotation group.The actiouis rotation of a point in the plane about the origin by an angle (1. The orbits are circles centered at the origin. The action is effective but not transitive. The stabilizer of every point in the plane is {eI, except the origin, for which the whole group is the stabilizer. Let M = SI, the unit circle, and G 80(2), the rotation group in two dimensions. The action is displacement of a point on the circle. There is only one orbit, the entire circle. The action is effective and transitive. The stabilizer of every point on the circle is {e}. Let M = G. a group, and let a (proper) subgroup S act on G by left multiplication. The orbits are right cosers Sg of the subgroup. The action is effective but not transitive. The stabilizer of every point in the group is {e}. Let M 1R U {oo}, the set of real numbers including "the point at infinity." Define an actionof 8 L (2, lit) on M by

=

=

a (c

realization ofa group

b).x=ax+c. d

bx +d

The reader may check that this is indeed a group action with a law of multiplication identical to the matrix multiplication, and that the action is transitive, but not effective. Let M be a set and H the group of transformations of M. Suppose that there is a homomorphism f : G --+ H from a group G into H. Then there is a natural action of G on M givenby g . m '" [f(g)](m). The homomorphism f is sometimes called a realization

ofG.

III

23.4 The Symmetric Group s, Because of its primary importance as the prototypical finite group, and because of its sigoificance in quantum statistics, the symmetric (or permutatiou) group is briefly discussed in this section. It is also used extensively in the theory of represeutation of the general linear group and its subgroups, A generic permutation n of n numbers is shown as

2 ".(2)

".(i)

(23.1)

Because the mapping is bijective, no two elements can have the same image, and ".(1), ".(2), ... , ".(n) exhaust all the elements in the set {iJi'=j'

665

23.4 THE SYMMETRIC GROUP 8"

e

"z

"Z e

"3

"6

"4 "S

"s "4

"6

"3

"3

"4

"s

"S e

"6

"3

"4

"S e

"4 "Z

"z

"3

"6

e

"z

e

"s

"6

"Z "4

"6

"3

Groupmultiplicatiou table for 83-

Table23.1

We can display the product "z 0 "1 of two penuutations using

"z 0 "I (i) ==

"z("I (i)). For instance, if 1 2 3 4) ( "1=3412 then the product

and

(I 2 3 4) "z=\243I'

"z 0 "1 takes I to 3, etc., because "z 0 "I (1) ea

(23.2) "Z("I (1))

=

"z(3) = 3, etc. We display "z 0 "I as

"z 0

"1 =

I 2 3 4) (3 1 2 4 .

23.4.1. Example. Let us constructthe multiplication table for 83- Deuote the elements of 83 as follows: e= "F

G ~ ;), G; ;),

G ~ ;), "s = G ~ ;), »z =

"3

= G;

"6 =

i),

G; i)-

Wegive only one sample evaluation of theentries andleavethe straightforward-c-but instructive---calculation of theother entries to thereader. Consider ft4 0 :n'5. andnote that "s(l) = 3 and"4(3) = 2; so''4o"s(l) = 2. Similarly, "40"S(2) = I and"4 0"S(3) = 3. Thus n'4 0 :lr5

2 3) = (I \2 1 3

= ]f2·

Theentire multiplication table is givenin Table 23.1.

III

Note that both the rows and columns of the group multiplication table include all elements of the group, and no element is repeated in a row or a column. This is because left-multiplication of elements of a group by a single fixed element of the group simply penuutes the group elements. Stated differently, the left multiplication map L g : G --> G, given by L g (x) = gx, is bijective, as the reader may verify. Because we are dealing with finite numbers, repeated application of a permutation to an integer in the set {i}i~1 eventually produces the initial integer. This leads to the following definition.

666

23. GROUP THEORY

cycles of symmetric group

23.4.2. Definition. Let:n: E Sn, i E {I, 2, ... , n}, and let r be the smallestpositive integer such that n" (i) = i. Then the set ofr distinct elements {:n: k (i)}~:b is called a cycle of:n: oflength r or an r-cycle generated by i. Start with 1 and app1y:n: to it repeatedly until you obtain 1 again. The collection of elements so obtained forms a cycle in which 1 is contained. Then we select a second number that is not in this cycle and apply :n: to it repeatedly until the original number is obtained again. Continuing in this way, we produce a set of disjoint cycles that exhausts all elements of {I, 2, ... , n}. 23.4.3. Proposition. Any permutation can be broken up into disjoint cycles. It is customary to write elements of each cycle in some specific order within parentheses starting with the first element, say i, on the left, then :n:(i) immediately to its right, followed by :n: 2(i), and so on. For example, the permutations :n:\ and :n:2 of Equation (23.2) and their product have the cycle structures rrj = (13)(24), :n:2 = (124)(3), and:n:2 o:n:\ = (132)(4), respectively.

23.4.4. Example. Let:n:\,:n:2

E Sg

be givenby

1 2 3 4 5 6 7 8)

:n:\

= (3

5

7

I

2

8

4

I

:n:2= ( 2

6 '

2 3 4 5 6 7 8) 5

6

8

I

7

4

3 .

Thereader mayverifythat 2 t

3 4 5 425

6 3

7 8

78)

and that

,,\ = (1374)(25)(68),

"2 = (125)(36748),

"20:n:\ = (16342)(5)(78).

In general, permutations do not commute. The product in reverseorder is

I

:n:\ o:n:2 = ( 5

2 3 4 5 6 7 8) 2

8

6

3

4

I

7 = (15387)(2)(46),

whichdiffersfrom.7l'2 osrj , However, notethatithasthesamecycle structure asJr2 01l'"I.in that 11II

cycles of equal lengthappearin both. This is a generalproperty of all permutations, cyclic permutations defined

23.4.5. Definition. If:n: E S« has a cycle of length r and all other cycles of:n: have only one element, then :n: is called a cyclic permutation of length r. It follows !hat:n:2 E Similarly,

S4

as defined earlier is a cyclic permutation of length 3.

1 2 3 4 5 6) 2 1 3 5 4

:n:= ( 6

is a cyclic permutation oflength 4 (verify this).

23.4 THE SYMMETRIC GROUP

transpositions defined

s;

667

23.4.6. Definition. A cyclic permutation oflength 2 is called a transposition. A transposition (Ij) simply switches i and j. 23.4.7. Example. Products of (not necessarily disjoint)cyclesmay be associated with a permutation whose action on i is obtained by starting with the first cycle (at the extreme right), locating thefirst occurrence of i, andkeeping track of what eachcycle does to it orits imageunder thepreceding cycle. Forexample, let 1fl E S6 be given as a product of cycles by rrj = (143)(24)(456). Tofind the permutation, we start with 1 aod followthe actionof the cycles on it, starting from the right. The firstand second cycles leave 1 alone, and the last cycle takes it to 4. Thus, 1fl (1) = 4. For 2 we note that the first cycle leaves it alone, the second cycle takesit to 4, aod the last cycle takes 4 to 3. Thus, "'\ (2) = 3. Similarly, "'\ (3) = 1, "'1 (4) = 5, "'\ (5) = 6, aod "'\ (6) = 2. Therefore,

1

"'\ = ( 4

2 3

3 4 1 5

5 6

6) 2 .

Wenote that 1l'1 is a cyclic permutation of length 6. It is left to the reader to show that the permutation "'2 "'2 = (13)(15)(12)(14) is cyclic: "'2 = (14253).

E

S5 given by the product 11II

The square of any transposition is the identity. Therefore, we can iuclude it iu any product of permutations without changing anythiug. 23.4.8. Proposition. An r-cycle (iI, [Z, ... , i,) can be decomposed into the product of r - 1 transpositions: (1\, tz, ... , i,) = (lli,)(i\i,-\)'" (1\;3)(1\;2).

Proof The proof iuvolves keepiug track of what happens to each symbol when acted upon by the RHS and the LHS and showiug that the two give the same result. This is left as an exercise for the reader. 0 parity ofa permutation defined

Although the decomposition of Proposition 23.4.8 is not mtique, it can be shown that the parity of the decomposition (whether the number of factors is even or odd) is unique. For iustance, it is easy to verify that

(1234) = (14)(13)(12) = (14) (34)(34) '(23) (12)(12)(23)(13)(12). '-v----' 1

'-v----' 1

That is, (1234) is written as a product of 3 or 9 transpositions, both of which are odd. We have already seen that any permutation can be written as a product of cycles. In addition, Proposition 23.4.8 says that these cycles can be further broken down iuto products of transpositions. This implies the followiug (see [Rotm 84, p.38]): parity ofa permutation is unique

23.4.9. Proposition. Any permutation can be decomposed as a product of transpositions. The parity ofthe decomposition is u n i q u e . '

668

23. GROUP THEORY

even and odd permutations

23.4.10. Definition. A permutation is even (odd) ifit can be expressed as a product ofan even (odd) number of transpositions. The parity of a permutation can be determined from its cycle structure and Proposition 23.4.8. The reader may verify that the mapping from Sn to the multiplicative group of {+I, -I} that assigns + I to even permutations and -I to odd permutations is a group homomorphism. It follows from the first isomorphism theorem (Theorem 23.2.16) that

23.4.11. Box. The set ofeven permutations, denoted by An ,forms a normal subgroup of Sn·

This homomorphism is usually denoted by

E.

We therefore define

ifn is even, ifnisodd.

_ {+I E(n ) = E" = -I

(23.3)

Sometimes 8(n) or 8" as well as sgn(n) is also used. The symbol, the definition of determinants, is closely related to E(n). In fact, En(l)Jl'(2) ...Jt(n)

Eili, .. .i,

used in

== En· -I

Suppose it, a E Sn, and note that? a(i) H i ~ neil ~ a 0 n(i), i.e., the composite a 0 no a- t of the three permutations takes a(i) to a 0 neil. This composite can be thought of as the permutation obtained by applying a to the two t 2 n) rows 0 f n = ( n(l) n(2) Jt(n) :

(a(l)

-1

a

0

n

0

a

=

a

0

n(l)

a (2) a

0

n(2)

a(n) ) a on(n) .

In particular, the cycles of a 0 n 0 a -1 are obtained by applying a to the symbols in the cycles of it . Since a is bijective, the cycles so obtained will remain disjoint. It follows that a 0 tt 0 a - t , a conjugate of n , has the same cycle structure as x itself. In fact, we have the following: 23.4.12. Theorem. Two permutations are conjugate same cycle structure.

if and

only

if they have

the

To find the distinct conjugacy classes of Sn, one has to construct distinct cycle structures of Sn. This in tum is equivalent to partitioning the numbers from I to n 7Recall from Chapter 0 that x

J..-" y means y = f(x).

23.5 PROBLEMS

669

into sets of various lengths. Let Vk be the number of k-cycles in a permutation. The cycle structure of this permutation is denoted by (I'", 2 V2 , ••• , n V, ) . Since the total number of symbols is n, we must have I:k~1 kVk = n. Defining Aj es I:k~j Vb we have (23.4) partition of n

Conjugacy classesof S, are related to the partition of n.

The splitting ofn into nonnegative integers (AI, A2, ... , An) as in Equation (23.4) is called a partition of n. There is a I-I correspondence between partitions of n and the cycle structure of Sn. We saw how Vk 's gave rise to A'S.Conversely, given a partition of n, we can construct a cycle structure by Vk = Ak - Ak+ I. For example, the partition (32000) of Ss corresponds to VI = 3 - 2 = I, V2 = 2 - 0 = 2, i.e., one I-cycle and two 2-cycles. One usually omits the zeros and writes (32) instead of (32000). When some of the A'Sare repeated, the number of occurrences is indicated by a power of the corresponding A; the partition is then written as (J.L~I, ,f.J-~r), where it is understood thatAl through AnI have the common value ILl, etc. For example, (3 21) corresponds to a partition of 7 with Al = 3, A2 = 3, and A3 = 1. The corresponding cycle structure is VI = 0, V2 = 2, and V3 = I, i.e., two 2-cycles and one 3-cycle. The partitions of length 0 are usually ignored. Since I: Ai = n, no confusion will arise as to which symmetric group the partition belongs to. Thus (32000) and (332000) are written as (32) and (322), and it is clear that (32) belongs to Ss and (322) to Ss.

tL?, ...

23.4.13. Example. Let us find the differeutcycle structuresof 84' This correspoods to different partitions of 4. We can take Al = 4 and the rest of the },,'s zero. This gives the partition(4). Next,welet a I = 3; then1.2 must be I, givingthepartition (31).WithAl = 2, A2 canbe either 2 or 1. In the latter case, ),,3 must be 1 as well, andwe obtaintwo partitions, (22) and (212). Finally, if Al = 1, all other nonzero A'smust be 1 as well (remember that Ak 2: Ak+I). Therefore, the last partitionis of the form (14 ) . Wesee that thereare 5 different partitions of 4. It follows that there are 5 different conjugacyclasses in 84. II

23.5 Problems 23.1. Let S be a subset of a group G. Show that S is a subgroup if and only if E S whenever a, b E S.

ab- I

23.2. Show that the intersection of two subgroups is a subgroup. 23.3. Let X be a subset of a group G. A word on X is either eo or an element w of G of the form

where Xi E X and e; = ± 1. Show that the set of all words on X is a subgroup of

G.

670

23. GROUP THEORY

23.4. Let la, b] denote the commutator of a and b. Show that (a) la, bj-I = [b, a], (b) [a,a] = eforalla E G, and (c) ab = la, b]ba. It is interestiog to compare these relations with the familiar

commutators of operators.

=

23.5. Show that if S is a subgroup, then S2 sa S S S, and t S t E S. More generally, T S = S if and only if T c S.

= S if and only if

23.6. Show that if S is a subgroup, then Sa = Sb if and only if ba:" E S and E Sand b-1a E S).

ab- I E S (as = bS if and only if a-Ib

23.7. Let S be a subgroup of G. Show that a

I>

b defined by ab- 1

E

S is an

equivalence relation. 23.8. Show that C G (x) is a subgroup of G. Let H be a subgroup of G and suppose x E H. Show that CH(X) is a subgroup of CG(x). 23.9. (a) Show that the only element a in a group with the property a 2 = a is the identity. (b) Now use ec * eo = eo to show that any homomorphism maps identity to identity, (c) Show that if f : G --> H is a homomorphism, then f(g-I) = [f(gJrl. 23.10. Establish a bijection between the set of right cosets and the set ofleft cosets of a subgroup. Hiot: Define a map that takes St to t- I S. Lagrange theorem

23.11. Let G be a finite group and S one of its subgroups. Convince yourself that the union of all right cosets of Sis G. Now use the fact that distioctright cosets are disjoiot and that they have the same cardinality to prove that the order of S divides the order of G. In fact, IGI = ISIIGISI, where IGISI is the number of cosets of S (also called the iodex of S io G). This is Lagrange's theorem. 23.12. Let f : G --> H be a homomorphism. Show that ¢ : G I ker f --> f (G) given by ¢(glker f]) sa ¢([ker f]g) = f(g) is an isomorphism. 23.13. Let G' denote the commutator subgroup of a group G. Show that G' is a normal subgroup of G and that GIG' is abelian. 23.14. Let M = IRU {oo}, and define an action of SL(2, IR) on M by

a b).x=ax+c. (c d bx +d Show that this is iodeed a group action with a law of multiplication identical to the matrix multiplication, and that the action is transitive, but not effective. 23.15. Show that two conjugacy classes are either disjoiot or identical. 23.16. Show that if all conjugacy classes of a group have only one element, the group must be abelian.

23.5 PROBLEMS

671

23.17. Consider a map from the conjugacy class of G containing x E G to the set of (left) cosets G/CG(x) given by q,(axa- t ) = aCG(x). Show that q, is a bijection. In particular, show that ICG(x)1 = IGI/IK~I where K~ is the class in G containing x and IK~I its order (see Problem 23.11). Use this result and Problems 23.8 and 23.11 to show that IHI/IK:I divides IGI/IK~I. 23.18. Show that R R. (0) R- 1 corresponds to a rotation of angle O. Hint: Consider the effect of rotation on the vectors in the plane perpendicular to and note that the rotated plane is perpendicular to f/ es Re.

e,

23.19. Let G act on M and let mo E M. Show that Gmo is the smallest invariant subset of M containing mo. 23.20. Suppose G is the direct product of HI and H2 and g = hlh2. Show that the factors h I and h2 are unique and that HI and H2 are normal. 23.21. Show that (g, h), (g', h') E GxH are conjugate ifandoulyifg is conjugate to g' and h is conjugate to h', Therefore, conjugacy classes of the direct product are obtained by pairing one conjugacy class from each factor. 23,22. Find the products ITI

ITI

0 IT2 and IT2 0 ITI

1 2 3 4 5 6) 4 6 5 1 2

of the two permutations

and

= (3

23.23. Find the inverses of the permutations

ITI

1 2 3 4 5 6 7 8) 5 7 I 2 8 4 6 '

= (3

and show directly that

(ITI 0 IT2)-1

=

1

Z

IT

0 ITil.

23.24. Find the inverse of each of the following permutations: IT2 =

(]

gg), IT3 =

(~B ~

W' and rra = <§ gW·

ITI -_

(1234) 3241'

23.25. Express each of the following products in terms of disjoint cycles. Assume that all permutations are in 87. (a) (123)(347)(456)(145).

(b) (34)(562)(273).

(c) (1345)(134)(13).

23.26. Express the following permutations as products of disjoint cycles, and determine which are cyclic. 2 3 4 5 6) (a) ( 1134562.

( b) (123456) 214563 .

2 3 4 5) (c) ( 113542.

. . Jr -- (12345678) 2327 . . E xpress th e permutation 2 4 1 3 6 8 7 5 as a pro"duet 0 f transpositions. Is the permutation even or odd?

672

23. GROUP THEORY

23.28. Express the following permutations as products of transpositions, and determine whether they are even or odd. (a)

G

1 2 3 (b)(4 1 7

(c)

(~

1 2 3 4 5 6 7) (d)(6 7 2 4 1 5 3 '

2 3 4 2 2.3

4 5

4 8

5 3

6 7 6 5

8) 2 .

23.29. Show that the product oftwo even or two odd permutations is always even, and the product of an even and an odd permutation is always odd. 23.30. Show that ic and n- 1 have the same parity (hoth even or both odd). 23.31. Find the number of distinct conjugacy classes of Ss and S6.

Additional Reading 1. Hamermesh, M. Group Theory and its Application to Physical Problems, Dover, 1989. The classic textbook on group theory written specifically for physicists. 2. Rottnan, J. An Introduction to the Theory ofGroups, 3rd ed., Allyn and Bacon, 1984. An excellent (formal, but readable) introduction to group theory with many examples and lots of explanations. 3. Wigner, E. Group Theory and its Application to the Quantum Mechanics of Atomic Spectra, Academic Press, 1959. Another classic written by the master of group theory himself.

24,

_

Group Representation Theory

Group action is extremely important in quantum mechanics. Suppose the Hamiltonian of a quantum system is invariant under a symmetry transformation of its independent parameters such as position, momentum, and time. This invariance will show up as certain properties of the solutions of the Schrodinger equation. Moreover, the very act of labeling quantum-mechanical states often involves groups and their actions. For example, labeling atomic states by eigenvalues of angular momentum assumes invariance of the Hamiltonian under the action of the rotation group (see Chapter 27) on the Hilbert space of the quantum-mechanical systemunder consideration.

24.1 Definitions and Examples In the language of group theory, we have the following situation. Put all the parameters Xl, , x p of the Hamiltonian H together to form a space, say IRP, and write H = H(XI, , xp ) H(x). A group of symmetry ofH is a group G whose action on IRP leaves Hunchanged,' i.e., H(x· g) = H(x). For example, a one-dimensional li2 d 2

=

harmonic oscillator, with H = - - 2 + imw2x2, has, among otherthings, 2mdx parity P (defined by Px = -x) as a symmetry. Thus, the group G = {e, P} is a group of symmetry of H. The Hamiltonian Hof a quantum-mechanical system is an operator in a Hilbert space, such as .c,2(IR3), the space of square-integrable functions. The important question is, What is the proper way of transporting the action of G from IRP

1It

will becomeclearshortly thatthe appropriate direction for the actionis fromtheright.

674

24. GROUP REPRESENTATION THEORY

to ,(,2(Jll3)? This is a relevant question because the solutions of the Schrodinger equation are, in general, functions of the parameters of the Hamiltonian, and as soch will be affected by the symmetry operation on the Hamiltonian. The answer is provided in the following definition. representation; carrier space and dimension a! a representation; taith!ul and identity representation

24.1.1. Definition. Let G be a group and ~ a Hilbert space. A representation of G on ~ is a homomorphism T : G --* G L (~). The representation is faithful if the homomorphism is 1-1. We often denote T(g) by Tg • ~ is called the carrier space of T. The trivial homomorphism T : G --* {1} is also called the identity representation. The dimension of~ is called the dimension ofthe representation

T. We do not want to distingnish between representations that differ only by isomorphic vector spaces, because otherwise we can generate an infinite set of representations that are trivially related to one another. A vector space isomorphism f: ~ --*~' induces a group isomorphism
GL(~).

This motivates the following definition. equivalent representations

24.1.2. Definition. Two representations T : G --* GL(~) and T' : G --* G L(~') are called equivalent if there exists an isomorphism f : ~ --* ~' such thatT~ = f 0 Tg 0 for all g E G.

r:'

24.1.3. Box. Any representation T : G --* GL(~) defines an action ofthe group G on the Hilbert space ~ by (g, 10)) sa Tg 10).

As we saw in Chapters 2 and 3, the transformation of an operator A under Tg would have to be defined by TgA(Tg)-I. For a Hamiltonian with a group of symmetry G, this leads to the identity Tg[H(x)](Tg)-l = H(x· g).

Similarly, the action of the group on a vector (function) in ,(,2(Ill3) is defined by (Tg 1/r)(x) sa 1/r(x, g),

(24.1)

where the parentheses around Tg 1/r designate it as a new function. One can show that if G acts on the independent variables of a function on the right as in Equation (24.1), then the vector space of such functions is the carrier space ofarepresentation of G. 10 fact,

24.1 DEFINITIONS AND EXAMPLES

675

where we have defined the new function


It follows from the last two equations that Tg",1/r = TgITg,1/r. Siuce this holds for arbitrary 1/r, we must have TgIg, = TglTez i.e., that T is a representation. When the action of a group is "naturally" from the left, such as the action of a matrix on a colmon vector, we replace x . g with g -1 . x. The reader can check that T : G --* GL(j{), given by Tg1/r(x) = 1/r(g-l . x), is iudeed a

representation. 24.1.4. Example. Let the Hamiltonian of the time-iudependent Schrodinger equation H11/r) = E 10/) be iuvariantunder the action of a group G. This means that TgHTgl = H Energyeigenstates can be labeled by eigenvalues ofthe symmetryoperators aswell.

=>

[H,Tgl = 0,

i.e., that H and Tg are simultaneously diagonalizable (Theorem 4.4.15). It follows that we can choose the energy eigenstates to be eigenstates ofT g as well, and we can label the states not only by the energy "quantum numbers"--eigenvalues of H-but also by the eigenvalues of Tg. For example, if the Hamiltonian is invariantunder the action of parity P, then we can choose the stales to be even, corresponding to parity eigenvalueof +1, or odd, corresponding to parity eigenvalueof -1. Similarly,if G is the rotation group, then the stales can be labeled by the eigenvalues of the rotation operators, which are, as we shall see, equivalent to the angular momentum operators discussed in Chapter 12. In crystallography and solid-state physics, the Hamiltonian of an (infinite) lattice is invariant under translation by an integer multiple of each so-called primitive lattice translation, the three noncoplanar vectors that define a primitive cell of the crystal. The preceding argument shows that the energy eigenstates can be taken to be the eigenstates of the translation operator as well. IIlII

matrix representations

It is common to choose a basis and represent all Tg'S in terms of matrices. Then one gets a matrix representation of the group G. 24.1.5. Example. Coosider the action of the 2D rotatioo group 80(2) (rotation about the z-axis) on JR.3:

x' r'=Rz(B)r

=>

= xcos9 -

ysin9,

y'=xsinB+ycosB,

z' w e. For a Hilbert space, also choose JR.3. Define the homomorphism T : G --+ GL(J[) to be the identity map, so that T(Rz(B)) ea Te = Rz(B). Tbe operator Te transforms the standard basis vectors of:J{ as

Teel = Te(l, 0, 0) = (cosB, sinB, 0) = cos Bel + sinBe2 + Oe3' Tee2 = Te(O, 1,0) = (- siuB, cos s, 0) = - siuBel + cosBe2 + Oe3, Tee3 = Te(O, 0, I) = (0,0, I) = Oel + Oe2 + e3.

676

24. GROUP REPRESENTATION THEORY

It follows that the matrix representation of 80(2) in the standard basisof j{ is

TO =

(

cos e s~e

- sioe case

o

0) O.

1

Note that 80(2) is an infinite group; its cardinality is determined by the "number" of 8~. g

24.1.6. Example. Let 83 act on lll.3 on the right by shnflling components:

For the carrier space, chooseR3 as well. Let T : 83 -+ GL(lR3) be given as follows: T(n:)

is the matrix that takes the column vectorx = consider" =

x;r(l)

Xl

(xz) X3

to

(X;r(2) ). X;r(3)

As a specificillustration,

(j j ~) and write Tn for T (x). Then

= (1,0,0)." = (0, 1,0) = e2, Tn(ez) = Tn (0, 1,0) = (0, 1,0)." = (0,0,1) = e3, Tn (e3) = Tn (0, 0, 1) = (0,0, 1)." = (1,0,0) = ej, Tn (e,) =Tn ( l , 0, 0)

whichgive riseto thematrix Tn =

0 0 1) (o 1

0 0 . 1 0

Thereader may construct the other five matrices of this representation andverifydirectly that it is indeed a (faithful) representation: Products andinverses of permutations are mapped

ontoproducts andinverses of the corresponding matrices.

Ill!II

The utility of a representation lies in our comfort with the structure of vector spaces. The climax of such comfort is the spectral decomposition theorems of (normal) operators on vector spaces of finite (Chapter 4) aud infinite (Chapter 16) dimensions. The operators Tg , relevaut to our present discussion, are, in general, neither normal nor simultaueously commuting. Therefore, the complete diagonalizability of all Tg'S is out of the question (unless the group happens to be abeliau). The best thing next to complete diagonalization is to see whether there are common invariant subspaces of the vector space 1i: carrying the representation. We already know how to construct (minimal) "invariaut" subsets of 1i:: these are precisely the orbits of the action of the group G on 1i:. The linearity of Tg'S gnarautees that the spau of each orbit is actually au invariaut subspace, aud that suchsubspaces arethesmallest invariant subspaces containing a givenvector. OUf aim is to find those minimal invariaut subspaces whose orthogonal complements are also invariaut. We encountered the same situation in Chapter 4 for a single operator.

24.1 DEFINITIONS AND EXAMPLES

reducible and irreducible representations

677

24.1.7. Definition. A representation T : G -+ G L (:Ji) is called reducible if there exist subspaces 11 and W of:Ji such that :Ji = 11 $ Wand both 11 and Ware invariant under all T8s. If no such subspaces exist,:Ji is said to be irreducible. In most cases of physical interest, where :Ji is a Hilbert space, W = 11.L. Then, in the language of Definition 4.2. I, a representation is reducible if a proper subspace of:Ji reduces all Tg's.

24.1.8. Example. Let 83 acton ~3 as inExample24.1.6.Forthecarrierspace %, choose

the space of functions on IR3, and for T, the homomorphism T : G --+ GL(9-C), given by t(x· g), for t E K Any t that is symmetric inx, y, z, snch as xyz, x + y +z, or xl + y2 + z2, defines a one-dimensional invariant subspace of Jf, To obtain another invariant subspace, consider VI (x, y, z) es xy and let {1fil?=l be as given in Example 23.4.1. Then, denoting T1l'i by Tj. the reader may check that

Tgt(x)

=

[Tltll(X, y, [T2tll(X, y, [T3tll(X, y, [T4tll(X, y, [T5tll(X, y, [T6tll(x, y,

z) = t l «x, y, z) . "1) z) = t l «x, y, z) . "2) z) = t l «x, y, z) . "3) z) = t l «x, y, z) . "4) z) = t l «x, y, z) . "5) z) = t l «x, y, z) . "6)

= t l (x, y, z) = xy = t l (x, y, z), = t l (y, x, z) = yx = t l (x, y, z), = t l (z, y, x) = zy == t2(X, y, z), = t l (x, z, y) = xz == t3(X, y, z), = t l (z, x, y) = zx = t3(x, y, z), = t l (y, z, x) = yz = t2(X, y, z).

This is clearly a three-dimensional invariant subspace of 9{ with 1/1'1, 1frz, and 1fr3 as a convenient basis, in which the first three permutations are represented by

Tl

= 01 01 0) 0 , T2 = (10 00 0) 1 , T3 = (01 01 0) 0 . ( 001

010

001

It is instructive for the reader to verify these relations and to find the three remaining matrices. II

24.1.9. Example. Let 83 act on ~3 as in Example24.1.6.For the catrier spaceofrepresentation, choose the subspace V of the J( of Example 24.1.8 spanned by the six functions x, y, z, xy, xz, and yz. For T, choose the same homomorphism as in Example 24.1.8 restrictedto V. It is clearthat the subspaces ti and W spanned,respectively, by the firstthree and the last three functions are invariant under 83, and that V 11 EEl W. It follows that the representation is reducible. The matrix form of this representation is found to be of the general form (~g), where B is one of the 6 matrices of Example 24.1.8. The matrix A,

=

correspondingto the three functions x, y, and z, can be found similarly.

II

Let Jf be a carrier space, finite- or infitrite-dimensional. For any vector 10), the reader may check that the span of (Tg 10) }gEG is an invariant subspace of :Ji.1f G is finite, this subspace is clearly fitrite-dimensional. The irreducible subspace containing 10), a subspace of the span of (T. la)}gEG, will also be finite-dimensional. Because of the arbitrariness of 10), it follows that every vector of :Ji lies in an irreducible snbspace, and that

678

24. GROUP REPRESENTATION THEORY

24.1.10. Box. All irreducible representations of a finite group are finitedimensional.

Due to the importance and convenience of unitary operators (for example, the fact that they leave the inner product invariant), it is desirable to be able to construct a unitary representation--or a representation that is equivalent to one-s-of groups. The following theorem ensures that this desire can be realized for finite groups. Allrepresentations are equivalent to unitary representations.

24.1.11. Theorem. Every representation ofafinite group G is equivalent to some unitary representation.

Proof We present the proof because of its simplicity and elegance. Let T be a representation of G. Consider the positive hermitian operator T == LXEG TtTx and note that Ting = ~)T(g)]t[T(x)]tT(x)T(g) xEG

= I)T(xg)]tT(xg) xeG

= L[T(y)]tT(y) = T, (24.2)

yeG

where we have used the fact that the sum over x and y == xg sweep through the entire group. Now let S = ..(f, and multiply both sides of Equation (24.2)-with S2 replacing T-by S-I on the left and by T;-IS-I on the right to obtain S-ITtg S ST-1S- 1 - g

=}

(STg S-I)t -- (STg S-I)-I

v g EG.

This shows that the representation T' defined by T~ sa STgS-1 for all g EGis unitary. D There is another convenience afforded by unitary representations:

24.1.12. Theorem. Let T : G ->- G L(:Ji) be a unitary representation and W an invariant subspace of:Ji. Then, W.l is also invariant. Proof Suppose la) E W.l. We need to show that Tg la) E W.l for all g E G. To this end. let Ib) E W. Then (bl Tg la) = «al Ti Ib))* = «al Tg1Ib))* = «al Tg-I Ib))* = 0, because Tg-I Ib) E W. It follows from this equality that Tg la) E W.l for all gEG. D The carrier space :Ji of a unitary representation is either irreducible or has an invariant subspace W, in which case we have :Ji = W Ell w-, where, by Theorem 24.1.12, W.l is also invariant. If W and w- are not irreducible, then they too can

24.1 DEFINITIONS AND EXAMPLES

679

be written as direct sums of invariant subspaces. Continuing this process, we can decompose J{ into irreducible invariant subspaces W(k) such that J{

=

W(I)

Ell W(2) Ell W(3) Ell ....

If the carrier space is finite-dimensional, which we assume from now on and for which we use the notationV, then the above direct sum is finite and we write p

V = W(l) Ell W(2) Ell ... Ell W(p)",

L

EllW(k).

(24.3)

k~1

One can think of W(k) as the carrier space of an (irreducible) representation. The homomorphism T(k) : G --+ GL(W(k») is simply the restriction of T to the subspace W(k), and we write r

Ts = T11) Ell T12) Ell ... Ell Tr) sa

L EllT1

k

)•

k~1

If we identify all equivalent irreducible representations and collect them together, we may rewrite the last equation as Tg

= mlT(l) « >

ffi

P

m2T(2)

«

ffi ••• ffi

m

T(P)

= "Ellm aT(a) g'

WPg-L-

>

(24.4)

a=l

where p is the number of inequivalent irreducible representations and m a are a positive integers giving the number of times an irreducible representation T1 ) and all its eqnivalents occur in a given representation. In terms of matrices, Tg will be represented in a block-diagonal form as

',{f:' jJ g

where some of the T1k) may be equivalent.

24.1.13. Example. A one-dimensional (and therefore irreducible) representation, definedfor all groups,is thetrivial (symmetric) representation T : G -+ iC givenby T (g) = t antisymmetric representation ofa permutation group

for all g E G. For the permutation group Sn, one can define another one-dimensional (thus irreducible) representation T : Sn ~ C, called the antisymmetric representation, given by T(lf) = +1 if rr is even, and T(lf) = -1 if If is odd. III

Given any (matrix) representation T of G, one can form the transpose inverse matrices (T~)-I, and complex conjugate matrices T;. The reader may check that each set of these matrices forms a representation of G. adjoint andcomplex conjugate representations

24.1.14. Definition. The set ofmatrices (T~) -1 and T; are called, respectively, the adjoint representation, denoted by '1, and the complex conjugaie representation, denoted by T*.

680

24. GROUP REPRESENTATION THEORY

24.2 Orthogonality Properties Homomorphisms preserve group structures. By studying a group that is more attuned to concrete manipulations, we gain insight into the structure of groups that are homomorphic to it. The group of invertible operators on a vector space, especially in their matrix representation, are particularly suited for such a study because of our familiarity with matrices and operators. The last section reduced this study to inequivalent irreducible representations. This section is devoted to a detailed study of such representations. Schur's lemma

24.2.1. Lemma. (Schur's lemma) Let T : G --+ GL(V) and T' : G --+ GL(V') be irreducible representations ofG. If A E I:(V, V') is such that

v g E G,

(24.5)

then either A is an isomorphism (i.e., T is equivalent to T'), or A = O. Proof Let la) E ker A. Then ATgla) =~Ala) =0

=}

Tgla} EkerA

v g EG.

'-,,-'

=0

It follows that ker A, a subspace of V, is invariant under T. Irreducibility of T implies that either ker A = V, or ker A = O. The first case asserts that A is the zero

linear transformation; the secondcase impliesthatA is injective. Similarly, let Ib) E A(V). Then Ib} = A [x) for some Ix) E V: ~ Ib} = ~A Ix) = ATg Ix) .

'---,---'

=}

Tg Ib} E A(V)

v g E G.

EA(I')

It follows that A(V), a subspace of V', is invariant under T'. Irreducibility of T' impliesthateitherA(V) = O,orA(V) = V'. The first case is consistent with the first conclusion drawn above: ker A = V. The second case asserts that A is surjective. Combining the two results, we conclude that A is either the zero operator or an isomorphism. D

Lemma 24.2.1 becomes extremely useful when we concentrate on a single irreducible representation, i.e., when T' = T. 24.2.2. Lemma. Let T : G --+ GL (V) be an irreducible representation of G. If A E I:(V) is such that ATg = TgAfor all g E G, then A = A1. Proof Replacing V' with V in Lemma 24.2.1, we conclude that A = 0 or A is an isomorphism of V. In the first case, A = O. In the second case, A must have a nonzero eigenvalue Aand at least one eigenvector (see Theorem 4.3.4). It follows that the operator A- A1 commutes with all Tg'S and it is not an isomorphism (why not?). Therefore, it must be the zero operator. D

24.2 ORTHOGONALITY PROPERTIES

681

We can immediately put this lemma to good use. If G is abelian, all operators {TxlxeG commute with oue another. Focusing on one of these operators, say Tg ,

noting that it commutes with all operators of the representation, and usiug Lemma 24.2.2, we conclude that Tg = A1. It follows that wheu Tg acts ou a vector, it gives a multiple of that vector. Therefore, it leaves any one-dimeusioual subspace of the carrier space invariant. Since this is true for all g E G, we have the following result.

24.2.3. Theorem. All irreducible representations of an abelian group are onedimensional.

This theorem is an immediate consequeuce of Schur's lemma, and is independent of the order of G. In particular, it holds for infinite groups, if Schur's lemma holds for those groups. One important class of infinite groups for which Schur's lemma holds is the Lie groups. Thus, all abelian lie groups have I-dimensional irreducible representations. We shall see later that the converse of Theorem 24.2.3 is also true for finite groups.

Issai Schur (1875-1945) wasoneof the mostbrilliaatmathematicians active in Germany during the first third of the twentieth century. He attended theGymnasium in Libau(now

Liepaja, Latvia) and then the University of Berlin, wherehe spent most cfhis scientific career from 1911 until1916.When he returned to Berlin, he was an assistant professor at Bonn. He became full professor at Berlin in 1919. Schur was forced to retire by the Nazi authorities in 1935 butwas able to emigrate to Palestine in 1939. He died there of a heart ailment several years later. Schur hadbeen a member of the Prussian Academy of SciencesbeforetheNazi purges. He married and had a son anddaughter. Schur's principal field was therepresentation theory of groups, founded alittlebefore 1900 by his teacher Frcbenlus. Schur seems to have completed this field shortly before

World War I, but he returned to the subject after 1925, when it became important for physics. Further developed by his student Richard Brauer, it is in our time experiencing anextraordinary growth through the opening of new questions. Schur's dissertation (1901) became fundamental totherepresentation theory of thegeneral linear group; in fact, English mathematicians havenamed certain of the functions appearing in the work"S-functions" in Schur's honor. In 1905Schur reestablished thetheory of group characters-the keystone of representation theory. Themostimportant tool involved is "Schur's lemma." Alongwith the representation of groups by integral linear substitutions, Schur was also the first to study representation by linear fractional substitutions, treating this moredifficult problem

ahnostcompletely in twoworks (1904, 1907). In 1906 Schurconsidered the fundamental problems that appear when an algebraic number field is taken as the domain; a number appearing in this connection is now called the Schur index. His works written after 1925 include a complete description of the rational and of the continuous representations of the general linear group; thefoundations of thiswork werein his dissertation.

682

24. GROUP REPRESENTATION THEORY

A lively interchange with many colleagues led Schur to contribute important memoirs to other areas of mathematics. Some of these were published as collaborations with other authors, although publications with dual authorship were almost unheard of at that time.

Here we simply indicatethe areas:pure group theory,matrices, algebraic equations,number theory, divergent series, integral equations, and function theory.

24.2.4. Example. Supposethatthe Hamiltonian H of a quantummechanical systemwith

n

Hilbert space has a group of symmetry with a representation T : G ~ GL(9-C). Then HTg = Tg H for all g E G. It follows that H = A1 if the representation is irreducible. Therefore,

24.2.5. Box. All vectors of each invariant irreducible subspace are eigenstates of the hamiltonian corresponding to the same eigenvalue, i.e., they all have the same energy. Therefore, the degeneracy of that energy state is at least as large as the dimension ofthe carrier space.

It is helpful to arrive at the statement above from a different perspective. Consider a vector Ix} in the eigenspace JV(j corresponding to the energy eigenvalue Ei. Since Tg and H commute, Tg [x} is also in JV(i. Therefore, an eigenspace of a Hamiltonian with a group of symmetry is invariant under all Tg for any representation T of that group. IT T is one of the II irreducible representations of G, say T(a) with dimension n a, then dimMi 2: n a.

Consider two irreducible representations TCa) and T(P) of a group G with carrier spaces w(a) and W(fJ), respectively. Let X be any operator in ,C(WCa), W(P»), and define A ea LT~a)XT~}I = L xeG

TCa)(x)XTCP)(x- 1) .

xeG

Then, we have T~a)A = LT(a)(g)TCa)(x)XTCP)(x-1)T(fJ)(g-I)T(P)(g) XEG

= LT(a)(gx)XT(P)«gx)-l) T(P)(g) = AT;fl. xEG

=A because this sum also covers all G

We are interested in the two cases where T(a) = T(P), and where T(a) is not equivalent to T(P). In the first case, Lemma 24.2.2 gives A = A1; in the second case, Lemma 24.2.1 gives A = O. Combining these two results and labeling the constant multiplying the unit operator by X, we can write LTCa)(g)XT(P)(g-l) = AX8ap1. gEG

(24.6)

24.2 ORTHOGONALITY PROPERTIES

683

The presence of the completely arbitrary operator Xindicates that Equation (24.6) is a powerful statement about-s-and a severe restriction on-the operators TCa) (g). This becomes more transparent if we select a basis, represent all operators by matrices, and for X, the matrix representation of X, choose a matrix whose only nonzero element is I and occurs at the lth row and mth column. Then Equation (24.6) becomes

L

T,~a)(g)T:!Y(g-l) = Alm8ap8ij,

geG

where Aim is a constant that can be evaluated as follows. Set j = i, Cl = fJ, and sum over i, TheRHS will give Aim Li 8ii = Almna, where n a is the dimension of the carrier space of r». For the LHS we get

LHS

=L geG

= L

LT~a)(g)T~I(g-l)

=L

(TCa)(g-I)T(a)(g))ml

gEG

i

T~7)(g-1 g) = L

geG

T~7)(e) = IG[8ml,

geG '-.---'

=(1)ml

where [GI is the order of the group. PUlling everything together, we obtain

"Ca) CP) -I [G[ LJ Til (g)Tmj (g ) = - 8mI8ap8ij,

nu

gEG

(24.7)

or (24.8) if the representation is unitary. Equations (24.7) and (24.8) depend on the basis chosen in which to express matrices. To eliminate this dependence, we first introduce the innportant concept of character. character 01 a representation; simple character, compound character

24.2.6. Definition. Let T : G --+ GL(V) be a representation ofthe group G. The character of this representation is the map X : G --+ C given by

X(g)

== trTg =

LTu(g), i

where T(g) is the matrix representation ofTg in any basis ofV. 1fT is irreducible, the character is called simple; otherwise, it is called compound. The character of the identity element in any representation can be calculated innmediately. Since a homomorphism maps identity onto identity, Te = 1. There-

fore, x(e) = tr(l) = dim V.

(24.9)

684

24. GROUP REPRESENTATION THEORY

Recall that two elements x, y E G belong to the same conjugacy class if there exist g E G such that x = gyg-I. This same relation holds for the operators representing the elements: Tx = Tg Ty Trl. Taking the trace of both sides, and noting that Trl = Tg1, one can show that

24.2.7. Box, All elements ofa group belonging to the same conjugacy class

have the same character. Setting i

= I and j = m in (24.7) and summing over i and i, we obtain

" ( . J (g)X" (OJ -I I G IL..,8ji " 8ij = -8.p I G IL..,8jj " = IGI8.p, L..,X (g ) = -8.p geG n« t.i n. j (24.10)

------=na

or

I: X(·J(g)X(PJ*(g) = IGI8.p

(24.11)

geG

if the representation is unitary. This equation suggests a useful interpretation: Characters can be thought of as vectors in a IG I-dimensional inner product space. According to Equation (24.11), the characters of inequivalent irreducible representations are orthogonal. In particular, since there cannot be more orthogonal vectors than the dimension of a vector space, we conclude that the number of irreducible inequivalent representations of a group cannot be more that the cardinality of that group. Actually, we can do better. Restricting ourselves to unitary representations and collecting all elements belonging to the same conjugacy class together, we

write

,

"L-CiXi(.J Xi(fJJ*

= IGI8af3

(24.12)

;=1

where i labels conjugacy classes, c; is the number of elements in the ith class, r is the number of classes in G, and Ix(·J) E
ki

X(g)

= X(g-I)

=}

Xi

= Xi',

where Kj' is the class consisting of all elements inverse to those of the class Ki. The equations involving characters of inverses of group elements can be written

24.3 ANALYSIS OF REPRESENTATIONS

685

in terms of the characters of the adjoint representation. For example, Equation (24.10) becomes

,

L

X(·)(g)X(Pl(g) = IGI8.p

=}

LCiXi·1x?) = IGI8.p.

(24.13)

i=l

gEG

Other relations can be obtained similarly.

24.3 Analysis of Representations We can use the results obtained in the last section to gain insight into a given representation. Take the trace of both sides of Equation (24.4) and write the result as p

X(g)=mlx(ll(g)+ ... +mpx(pl(g)

== Lm.x(·)(g);

(24.14)

Cl!=1

i.e., a compound character is a linear combination of simple characters with nonnegative integer coefficients. Furthermore, the orthogonality of simple characters gives (24.15)

yielding the number of times the irreducible representation T(·) occurs in the representation T. Another useful relation will be obtained if we multiply Equation (24.14) by its complex conjugate and sum over g; the result is

L

Ix(g)1 2 = L X(g)X*(g) = L Lm.X(·l(g) LmpX(Pl*(g)

geG

geG

geG a

f3

= Lm.mp L X(·)(g)X(PJ*(g) = IGI Lm~. "p gEG a

(24.16)

=IGI'.~

criterion for irreducibility

In particular, if T is irreducible, all m. are zero except for one, which is unity. We therefore obtain the criterion for irreducibility:

L geG

,

Ix(g)1

2

= L cilxil 2 = IGI

if T is irreducible.

(24.17)

i=l

For groups of low order and representations of small dimensions, Equation (24.16) becomes a powerful tool for testing the irreducibility of the representation.

686

24. GROUP REPRESENTATION THEORY

24.3.1. Example. Let a = 83 and consider the representation of Exampte 24.1.8.The characters of the first three elements of thisrepresentation are easily calculated: Xl

X2 = trT2

= tr Tj = 3,

= I,

X3

ee

tr Tj

= I.

Similarly, one can ohtain X4 = I, X5 = 0, and X6 = O. Substituting this in Eqnation(24.16) yietds

L geG

Ix(g)1

2

6

=

L

Ixjl2

= 32 + 12 + 12 + 12 + 02 + 02 = 12.

j=l

Comparing this with the RHS of (24.16) with 101 = 6 yields I:am~ = 2. This restricts thenonzero a's to two, say a = 1 and ex = 2. Moreover, ml and mz can be only 1. Thus, therepresentation of Example 24.1.8 is reducible, andthere arepreciselytwo inequivalent irreducible representations in it, eachoccurring once. Wecanactually findtheinvariant subspaces corresponding to thetwoirreducible representations revealed above. Thefirst is easyto guess.Justtaking thesumof the three functions tlo t2, and t3 gives a one-dimensional invatiantsnbspace; so, let ¢I sa tl + t2 + t3, andnote that the space WI spanned by cPt is invariant. The second is harder to discover. However, if we assume that 1/11, 1/12. and 1/13 areorthonormal, then usingthe Gram-Schmidt process, we can find the other two functions orthogonal to cPt (but not orthogonal to each other!). These are

The reader is nrged to convince himselflherselfthat the subspace W(2) spannedby ¢2 and is the complement of W(l) [i.e., V = W(l) Ell W(2)j and that it is invatiant under all

¢3

Tg's.

regular representation

II

A very useful representation can be conslrncted as follows. Let G = {gj It=l' and recall that left multiplication of elements of G by a fixed element gi is a permutation of (gl' g2, ... ,gm). Denote this permutation by :J(i. Now define a representation R ; G ---> GL(IRm ) , called the regular representation, by

That this is indeed a representation is left as a problem for the reader. One can obtain a matrix representation of R by choosing the standard basis {ej lj'=l ofIRm and noting that Rg;ej = en,-l(j), From such a matrix representation itfollows that all characters X R of the regular representations are zero except for the identity, whose character is xR(e) = m [see Equation (24.9)]. Now use Equation (24.14) for g = e and for the regular representation to obtain m = I:~=1 mana where n a is the dimension of the a-th irreducible representation. We can find m a by using Equation (24.15) and noting that only g = e contributes to the sum:

24.4 GROUP ALGEBRA

687

In words,

24.3.2. Box. The number of times an irreducible representation occurs in the regular representation is equal to the dimension ofthat irreducible representation.

We therefore obtain the important relations p

p

xl = IGI 8i1 =

Lnaxia)

and

a=l

IGI= Ln~,

(24.18)

a=l

where we have assumed that the first conjugacy class is that of the identity. For finite groups of small order, the second equation can be very useful in obtaining the dimensions of irreducible representations. 24.3.3. Example. A group of order 2 or 3 has only one-dimensional inequivalent irreducible representations, because the only way that Equation (24.18) can be satisfied for IG I = 2 or 3 is for all nO!'s to be 1. A group of order 4 can have either 4 one-dimensional or one 2-dimensional inequivalent irreducible representations. The symmetric group 83. being of order 6, can have 6 one-dimensional, or 2 one-dimensional and one 2-dimensional inequivalent irreducible representations. We shall see later that if all inequivalent irreducible representations of a group are one-dimensional, then the group must be abelian. Thus, the

first possibility for 83 must be excluded.

III

24.4 Group Algebra

group algebra defined

Think of group elements as (linearly independent) vectors. In fact, given any set, one can generate a vector space by taking linear combinations of the elements of the set assumed to form a basis. In the case ofgroups one gets a bonus: The product already defined on the basis (group elements) can be extended by linearity to all elements of the vector space to turn it into an algebra called the group algebra. For G = {gj lj'=to a typical element of the group algebra is a = L:~=t aigi. One can add two vectors as usual. But the product of two vectors is also defined:

where Ok is a sum involving example.

a,

and bi- The best way to leam this is to see an

688

24. GROUP REPRESENTATION THEORY

24.4.1. Example. LetG = S3 and consider a = 211"t-311"3+11"5 andb = 11"2-211"4+311"6' Then,usingTable 23.1, we obtain ab = (211"1 - 311"3 + 11"5)(11"2 - 211"4 + 311"6)

= 21l'I1l'z -

41l'1n'4

- 91l"31l'6

+ 61l'pr6 -

+ 1r51C2 -

3n'31l'2

2.7Z'57l'4

+ 6Jr31f"4

+ 31l'51r6

=~-~+~-~+~-~+~-~+~

= 311"1 - 711"2 - 211"3 - 311"4 + 611"5 + 311"6.

III

Let A be any algebra. As a finite-dimensional vector space, we can always find two proper subspaces ,c1 and,c2 such that A is a direct sum Of,c1 and ,c2. We write this as 2 A = ,c1 ,c2. If we demand that this sum be invariant under left (right) algebra multiplication, then it is clear that ,c1 and,c2 must be left (right) ideals, in whichcase we write

+

Now assume thatA has an identity 1,which as a vector in A can be decomposed uniquely as (24.19)

Y a EA.

Multiplying these two equations, we obtain

The uniqueness of the decomposition of a now implies that

So, if a

E

,c1, then a2 = 0 and

with a similar result for a E ,c2. Since 11 E ,c1 and 12 E ,c2, we have (24.20) idempotent element of algebras

An element a E A that satisfies a 2 = a is called an idempotent.f Thus, 11 and 12 are idempotents. Furthermore, they generate L, and ,c2, respectively; i.e., ,c1 = All and,c2 = A12 (see the last section of Chapter 1).

2Thereasonfor changing to this new notation is to reserveEEl for the following. 3Inthe algebra of operators, we called suchan elementa projection operator.

24.4 GROUP ALGEBRA

689

24.4.1 Group Algebra and Representations Group algebra is very useful for the construction and analysis of representations of groups. In fact, we have already used a similar approach in the construction of the regular representation. Instead of IR"' used before, use the m-dimensional vector space A, the group algebra. Then left-multiplication by a group element g can be identified with T~R) , the operators of the regular representation, and the invariant subspaces of A become the left ideals of A, and we can write

resolution ofthe identity

Moreover, since the identity element of the group is the identity element of the algebra as well, the argument at the end of the last subsection gives the resolution"

e = el essentially idempotent elements

+ ... + er

(24.21)

It is clear that if a 2 = ua, then a/a will be idempotent. So, we can essentially ignore the constant a, which is why a is called essentially idempotent. Now consider the element of the gronp algebra

p=I>

(24.22)

XEG

and note that g P =

p

2=

gx = P. It follows that

LgLx= L L gx= L geG

primitive idempotents

LXEG

xeG

geG xeG

P=

IGIP.

geG

So, P is essentially idempotent. Furthermore, the reader may verify that the ideal generated by P is one-dimensional. An idempotent that cannot be resolved into other idempotents satisfying Equation (24.21) is called a primitive idempotent. The reader may check that the following holds. 24.4.2. Proposition. A left ideal is minimal if and only if it is generated by a primitive idempotent. Let us now apply the notion of the group algebra to derive further relations aroong characters. Denote the elements of the ith class KI of G by and construct the element of the group algebra e, '" such quantities ci

Ki/(j

{x?)lf::l Lf::1 x?J. If in the product oftwo

Cj

" 'L..J " Xl(I) Xm (j) , = 'L...J

1=1 m=l

4Inthe algebra of operators, we calledelements satisfying theserelations orthogonal projection operators.

(24.23)

690

24. GROUP REPRESENTATION THEORY

x?)x;;p

sa Y E G, is in a certain conjugacy class, then the rest of that class can be obtained by taking all conjugates of y, i.e., elements of G that can be written as

It follows that if one member of a class appears in the double sum of Equation (24.23), all members will appear there. The reader may check that if y occurs k times in the double sum, then all members of the class of y occur k times as well. Collecting all such members together, we can write r

(24.24)

KiKj = LCijIKl, 1~1

where Cijl arepositive integers. Now consider the ath irreducible representation, and add all operators corresponding to a given class: r i(a) =- "r(a) L.J g

r

~ --r

r(a)r(a) - " c .. r(a) t

j

-

L."

geKj

ijl I

'

(24.25)

[=1

where the second equation follows from the same sort of argument used above to establish Equation (24.24). One can show that ria) commutes with all r1a). Therefore, by Schur's lemma, ria) = Aia)1, and the second equation in (24.25) becomes

, Aia)Aja) = Lc;jIA!a)

(24.26)

1~1

Taking the characters of both sides of ria) = Aia) 1 and using the first equation in (24.25), noting that all elements of a class have the same character, we get

Substituting this in Equation (24.26), we obtain

, CiCjxtt)xJct) = n« LCij1CiXl(Cl).

(24.27)

1=1

This is another equation that is useful for computing characters. Note that this equation connects the purely group properties (c; 's and Cijl 's) with the properties

24.4 GROUP ALGEBRA

691

of the representation (xia),s and na). Summing Equation (24.27) over a and using the first equation in (24.18), we get p

CiC)

r

p

'L-Xi " (a) Xj(a) = 'L-CijlCJ " ct=l

[=1

'L-naXI " (a)

= CfjllGI

a=l '---,,-'

=IGlolI by (24.18)

because CI = 1 (there is only one element in the class of the identity). Problem 24.12 shows that Cijl = Ci[,i' j where Kjl is the class consisting of inverses of elements of Ki, It then follows that (24.28) For a unitary representation, x5a) = xia)*, so Equation (24.28) becomes (24.29) where IXi) E iC P is a p-dimensional vector with components {Xia)}~=I' This equation can also be written in terms of group elements rather than classes. Since a) xi = X(a)(x) for any x E Ki, we have

i:

a=l

X(a)(x)X(a)*(y) =

1~~18(K:, K~),

(24.30)

x

where K: is the conjugacy class of G containing x, elements, and

IK:I

is the number of its

Equation (24.29) shows that the r p-dimensional vectors xia) are mutually orthogonal; therefore, r :::: p. The statement after Equation (24.12) was thatr 2: p. We thus have the following: 24.4.3. Theorem. The number of inequivalent irreducible representations of a finite group is equal to the number of conjugacy classes in the group.

character table of a finite group

It is convenient to summarize our result in a square table with rows labeled by the irreducible representation and columns labeled by the conjngacy classes of G. a) Then on the ath row and i th colunm we list xi , and we get Table 24.1, called the character tahle of G. Note that Ci, the order of Ki, is written as a left superscript.

692

24. GROUP REPRESENTATION THEORY

«x, T(I)

«x,

(I)

(I)

X~2)

Xi

T(2)

Xl (2) Xl

T(a)

(a) XI

(c)

X2

xi

T(r)

Table 24.1

(I)

c2K2

(,)

XI

X2

(,)

X2

(2)

Xi a

)

(,.)

Xi

Cr

K, (I)

X,

(2)

X,

(a) X, (r) X,

A lypicalcharacter table.

Character tables have the properly that any two of their rows are orthogonal in the sense of Eqnation (24.12), and any two of their columns are orthogonal in the sense of Eqnation (24.29). If all inequivalent irreducible representations of a group G have dimension one, then there will be IGI of them [by Equatiou (24.18)]. Heuce, there will be IGI conjugacy classes; i.e., each class consists of a single element. By Problem 23.16, the group must be abelian. Combining this with Theorem 24.2.3, we have the following theorem. 24.4.4. Theorem. A finite group is abelian if and only if all its inequivalent irreducible representations are one-dimensional.

24.5

Relationship of Characters to Those of a Subgroup

Let H be a subgroup of G. Denote by Kf! and Kf/ the H-class containing h E H and the G-class containing g, respectively. Let d j and ci be the number ofelements in the jth H -class and ith G-class, respectively. Any represeutation of G defines a representation of H by restriction. An irreducible representation of G may be reducible as a represeutation of H. This is because although the subspace w(a) of the carrier space that is irreducible under G is the smallest such subspace containing a given vector, it is possible to generate a smaller subspace by applying a subset of the operators Tg corresponding to those g's that belong to H. It follows that T(a)(h) = Lma,,-t("-)(h), o

h e H,

(24.31)

where m a,,- are nonnegative integers as in Equation (24.14) and t("-) are irreducible representations of H. If X (a) and ~(,,-) denote the characters ofirredncible representations of G and H, respectively, then the equivalent equation for the characters

24.5 RELATIONSHIP OF CHARACTERS TO THOSE OF A SUBGROUP

693

is XCa\h) = Lrna"~C"\h),

h e H.

(24.32)

o

Multiply both sides by ~CK)*(h), sum over h E H, aud take the complex conjugate at the end. Then by the orthogonality relation (24.11), applied to H, we obtain rnaK =

I~I

L

X(a)*(h)~CK)(h).

(24.33)

heR

Now multiply both sides of Equation (24.33) by X(a)(g), sum over ct, aud use Equation (24.30) to obtain "rn L.., a

cac

X(a)(g) =

IGI

IHI IKGI g

"8(K G KG).CK)(h). L..,

heH

h'

g

5

(24.34)

The sum on the right cau be trausformed into a snm over conjugacy classes of H. Then Equation (24.34) becomes "

(a) _

L..,rnaKXi a

-

JQL .• (K) IHI . "L.., d j5j , Cl

i = 1,2, ... , r,

(24.35)

j

where the sum on the LHS is over irreducible representations of G, aud on the RHS it is over those H -classes j that lie in the ith G-class. Note that the coefficients IGldj/(IHlci) are integers by Problem 23.17. Equations (24.34) aud (24.35) are useful for obtaining characters of G when those of a subgroup H are known. The general procedure is to note that the RHS of these equations are completely determined by the structure of the group G aud the characters of H. Varying i, the RHS of (24.35) determines the r components of a (compound) character IJ/r), which, by the LHS, can be written as a linear combination of characters of G:

IJ/r)

r sa Lrna Ix Ca»), «=1

(24.36)

where we have suppressed the irrelevaut subscript 1<:. If we know some of the Ix(a»)'s, we may be able to determine the rest by taking successive inner products to find the integers rna, aud subtracting each irreducible factor of the sum from the LHS. We illustrate this procedure for S« in the following example. 24.5.1. Example. Let K 1 = (12) and K2 = (2) for S2 (see Section 23.4 for notation). Example 24.1.13 showed that we canconstruct two irreducible representations for any Sn, thesymmetric and theantisymmetric representations. Thereader mayverifythat these two representations areinequivalent. Since thenumber of inequivalent irreducible representations is equal to thenumber of classes in a group, we have all the information needed to

694

24. GROUP REPRESENTATION THEORY

1 1 Table 24.2

1 -1

Character table for 82.

T(l) T(2) T(3)

Table 24.3

IKj

3K2

2K3

I 1 ?

I -1 ?

1 I ?

Partially filled character table for 83.

construct the character table for 82. Table 24.2 shows this character table. We want to use the 82 character table to construct the character table for 83. With our knowledge of the symmetric and the antisymmetric representations, we can partially fill in the 83 character table. LetKj = (1 3 ) , K2 = (2, I), aod K3 = (3) aodnote that cj = I, C2 = 3, aod C3 = 2. Then we obtain Table 24.3. To complete the table, we start with K 1, and write the RHS of Eqnation (24.35) as

=

>/Ii

=

~ Ld·«I) = ~ L«I) 2ei

J J

j

Ci

j

J

== 1 for the two classes of 82. The sum on the RHS is over Sz-cIasses that are

because d j

inside the ith S3-class. For i = 1, only the the first Sz-class contributes. Noting that are the entries of Table 24.2, we get

>/It

gy)

3 (I) 3 = -
CJ

I

Similarly,

3 (1) 3 >/12 = -<2 = - . I = I ~

3

aod

3 >/13 = - . 0 = O. ~

The second equation follows from the fact that there are no classes of 82 inside the third class of 83. Eqnation (24.36) now gives

1>/1)

=

(i)o = i:>.

Ix(·»).

a=l

We can find the number of times Ix(l)) occurs in this compound character by taking the inner product:

(x(l) I>/I) =

r

L m. (x(l) Ix(·») = mIlGI = 6ml· a=l

24.6 IRREDUCIBLE BASIS FUNCTIONS

T(l) T(2) T(3)

1Kl

3K2

2K3

1 1 2

1 -1 0

1 1 -1

695

Table 24.4 Complete character tablefor 83.

But

,

(x(l) 11/1) = I:; c; Xil)1/Ii = I . I ·3 +

3· I . 1+2· I ·0 = 6.

i=l

These two equations showthat m 1 = 1. So,

Subtracting the columnvectors,we get a new character:

Takingthe inner product with 1X (2» yields m2 = O. It follows that 11/1') is a simple cbaracter. In fact,

I:; c;l1/l[1 2 =

I . 22 + 3 . 02 + 2 . (_1)2 = 6,

aod the criterion of irreducibility, Equation (24.17), is satisfied. Wecannow finish up Table 24.3 to obtain Table 24.4, whichis the completecharacter table for 53. III

24.6 Irreducible Basis Functions We have studied the operators T. and their characters representing group elements in rather extensive detail. Let us now tum our attention to the carrier space itself. In particular, we want to concentrate on the basis functions of the irreducible

representations. We choose "functions," rather than vectors, because of their use in quantum mechanics as discussed at the beginning of this chapter. Let {Iti a») be a set ofbasis functions forW(a), the ath invariant irreducible subspace. Invariance of wCa) implies that

1:':1

T.

Iti·») =

n.

L Tj~·) (g) Itja»), j~1

696

24. GROUP REPRESENTATION THEORY

where Tj~a) (g) are elements of the matrix functions beionging to the ith row ofthe ath irreducibie representation

T1a ) representing g E G.

24.6.1. Definition. Afunction (or vector) I


)};:!

irreducible representation ofG with matrices (Tj~a) (g») and n a - latherfunctions {1
T. 1


(24.37)

j~!

Fnnctions that belong to rows of irreducible representations have some remarkable properties. Let 11ft/aJ) and l
(1ft/a ) I
= L

np

L

Tz~a)*(g)T~J (g)

(1ftiaJI
[=1 m=l

Summing this equation over g yields IGI (1ft/aJI
RHS = L

nfJ,

L

.

,

Tz~aJ*(g)T~)(g) (1ftiaJI
L

[=1 m=lgeG

-

~BapUI] e. ~(.I.
[=1

where we have made use of Equation (24.8). Therefore,

(1ft/a) I
t (1fti

a

I

)


(24.38)

1=1

This shows that functions belonging to different irreducible representations are orthogonal. We should expect this, because in our construction of invariant irreducible subspaces, we kept dividing the whole space into orthogonal complements. What is surprising is thatfunctions transforming according to different rows ofan irreducible representation are orthogonal. We had no control over this property! It is a consequence of Equation (24.37). Another surprise is the independence of the inner product from i: If we let i = j and a = fi on both sides of (24.38), we obtain (24.39)

24.6 IRREDUCIBLE BASIS FUNCTIONS

697

which indicates that the inner prodnct on the LHS is independent of i, symmetry and the quantum mechanical perturbation theory; lilting of degeneracy

24.6.2. Example. The quantum-mechanical perturhation theory starts with a known Hamiltonian Ho witheigenvalues Ej andthe corresponding eigenstates IEj}. Subsequently, a (small) perturbing "potential" V is added to the Hamiltonian, and the eigenvalues and eigenstates of the new Hamiltonian H = Ho + V are sought. One can draw important conclusions about the eigenvalues and eigenstates of the total Hamiltonian by symmetry arguments. Supposethe symmetry group of HO is G, and that of H is H, whichhas to he a subgroup of G. In most cases, the eigenspaces of HO are irreducible carrier spaces of G, i.e., their basis vectors transform according to the rows of irreducible representations of G. If H is a proper suhgroupof G, then the eigenspacesof Ho will split accordingto Equation (24.31). We say thatsome of the degeneracy is lifted because of the perturbation V. The nature of the split, i.e., the number and the dimensionality of the vector spaces into which a given eigenspacesplits,can be obtainedhy the charactersof G and H and Equation (24.32). The original eigenspaces are represented on an energy diagram with a line corresponding to each eigenspace.The split of the eigenspace into knew subspaces is thenindicated by the branching of the old line into k new lines. To the lowest approximation-first-order perturbation theory-the magnitude of the split, i.e., the difference between the eigenvalues of HO and those of H, is given by [see Equation (21.57)] the expectation value (oPi·) IV loP;·»), where loPi·») helongs to the ith row of the «th irreducihle representation, and loP]"») to its jth row (i # j). Only if this expectation value is nonzero will a split occur. This, in turn, depends on the symmetry of V: If V is at least as symmetric as Ho (corresponding to G = H), then (¢ja) V ItP)a)} = 0, and no splitting occurs (Problem 24.17). If, on the other hand, V is less symmetric than Ho (corresponding to H c G), then V loP;·») will not belong to the jth row of the cth

I

irreducIble representation, and in general, (oPi·) IV loP;·»)

# O.

III

We had decomposed the carrier space V of a representation into invatiant irreducible subspaces WC.). The argument above shows that each WC.) has a basis consisting of the "rows" of the irreducible representations. Corresponding to such a basis, there is a set of projection operators with the property = 1 (Chapter 4). Our aim is to find an expression for these operators, which have the Itia») = Iti·»). We start with Eqnation (24.37), multiply defining property

pl·)

L.,i pla)

pl·)

both sides of it by T/~). (g), sum over g E G, and use Equation (24.8) to obtain

L Tl~)·(g)Tg Iti·»)

n.

=

Itya») L

j=l

gEG

= Let fJ = becomes

L

01,

m

-IGI ~ LJ na j=l

T/~)·(g)Tj~a)(g)

geG

It j(a») 8tj8mi 8. p =IGI- It 1(.») 8mi 8. p.

= I = i, and multiply both sides

na

by n a / 1G I. Then this equation

698

24. GROUP REPRESENTATION THEORY

projection operator onto the i1h row of the eth irreducible representation

which suggests the identification p(.) I

= ~ "" r.(.)*( )T IGI LJ /I g s

(24.40)

geG

with the properties

(24.41) a

where l4>i ») is the projectiou of 14» aloug the ith row of the ath irreducible representation. We are also interested in the projection operator that projects outo the irreducible subspace W(·) . Such an operator is obtained by summing over i. We thus obtain

pi·)

projection operator onto the ath irreducible representation

n

p(.)

n

na

L L r,\a)*(g) T 1 1 ,~,

= ;

geG

.

g

= ;

L X (·)*(g)T

1 1geG

g

(24.42)

=x(oo*(g)

and

(24.43) where 14>(·» is the projection of 14>} onto the ath irreducible invariant subspace. These formulas are extremely useful in identifying the irreducible subspaces of a given carrier space: Start with a basis {Ia,)} of the carrier space, apply p(.) to all basis vectors, and collect all the linearly independent vectors of the form p(.) la,). These vectors form a basis of the oth irreducible representation. The following example illustrates this point.

24.6.3. Example. Consider the representation of 83 given in Example 24.1.8, where the carrier space is the span of the three functions It,) = xy, It2) = yz, and It3) = xz. We refer to the character table for 83 (Table 24.4) and nse Equation (24.42) to obtain p(l) =

~(T, +T2 +T3 +T4 + Ts +T6),

p(2) =

~(T, -T2 -T3 -T4+TS +T6),

p(3) =

~(2T, -

6

6

TS - T6),

where,asinExample24.1.8.wehaveusedthenotationr, forT1l'j,andtheresultnl = nz = 1 and n3 = 2 obtained from Eqnation (24.18), Theorem 24.4.4, and the fact that 83 is nonabelian. To get the first irreducible subspace of this representation, we apply p(l) to ItI). Since this subspace is one-dimensional, the procedure will give a basis for it if the vector so

24.7 TENSOR PROOUCT OF REPRESENTATIONS

699

obtained is nonzero: p(l)

I'h) =

~(T1 +T2 + T3 + T4 +TS +T6) I'h) I

I

= "6(1"'1) + 1"'1) + 1"'2) + 1"'3) + 1"'3) + 1"'2)) = 3(1"'1) + 1"'2) + 1"'3))'

This is a basisfor thecarrier spaceof the irreducible identity representation. Forthe secondirreducible representation, we get

Similarly, p(2) 1"'2)

= 0 and p(2) 1"'3) = O. This means that T(2)

is not inclndedin the

representation we are working with. We shouldhave expected this, because if this onedimensional irreducible representation were included, it would force the last irreducible representation to be one-dimensional as well [see Equation (24.18)], and, by Theorem 24.4.4, the gronp S3 to be abelian! The lastirreducible representation is obtained similarly. Wehave

These two vectors are linearly independent. Therefore, they form a basis for the last irreducible representation. The reader may check that p(3) It3} is a linear combination of p(3) 1"'1) and p(3) 1"'2)' l1li

24.7 Tensor Product of Representations A simple quantum mechanical system possessing a group of symmetry is described by vectors that transform irreducibly (or according to a row of an irreducible representation). For example, a rotationally invariant system can be described by an eigenstate of angnlar momentum, the generator of rotation. s These eigenstates transform as rows of irreducible representations of the rotation group. At a more fundarnentallevel, the very concept of a particle or field is thought of as states that transform irreducibly under the fundamental group of spacetime, the Poincare group. Often these irreducible states are "combined" to form new states. For example, the state of two (noninteracting) particles is described by a two-particle state, labeled by the combined eigenvalues of the two sets of operators that describe each particle separately. In the case of angular momentum, the single-particle states may be labeled as Iii, mil for i = 1,2. Then the combined state will be labeled as 111, m1; 12, m2), and one can define an action of the rotation group on 5Chapter 27 will make explicitthe connection betweengroups ~d theirgenerators.

700

24. GROUP REPRESENTATION THEORY

the vector space spannedby these combined states to construct a representation. We now describe the way in which this is done. Snppose that T : G -+ GL(V) and S : G -+ GL(W) are two representations ofa gronp G. Let V<8lWbe the tensor product of V and W (see Example 1.3.19). Now define an action of the group G on V <8l W via the representation T <8l S : G -+ GL(V <8l W) givenby T <8l S(g)(lv), Iw» = (T(g) Iv), S(g) Iw».

We note that T <8l S(gjgz)(lv), Iw» = (T(gjgz) Iv), S(gjgz) Iw}) = (T(gj)T(gz) Iv), S(gj)S(gz) Iw})

= T <8l S(gj)(T(gz) [u}, S(gz) Iw» = T Kronecker product representation

e S(gj)T e S(gz)(lv}, Iw».

It followsthat T <8l S is indeeda representation, called the tensor prodnct or direct product or Kronecker product representation. It is common, especially in the physicsliterature,to write [u, wi, or simply Ivw) for (lu) , Iw}), and T S for T <8l S. If we choose the orthonormal bases Ilvi}} for V and Ilwa}} for W, and define an

inner product on V <8l W by (v, wi v', w') ea (vi v') (wi w'),

we obtain a matrix representationof the gronp with matrix elements givenby (T <8l S)ia,jb(g)

== (Vi, wal T e S(g) IVj, Wb} = (vii T(g) IVj) (wal S(g) IWb)

character of a product representation isa product of characters

== T;j(g)Sab(g).

Note that the rows and columns of this matrix are distinguishedby doubleindices. If the matrix T is m x m and S is n x n, then the matrix T <8l S is (mn) x (mn). The character of the tensor product representationis XT®s(g)

= L(T <8l S);a,ia(g) = L t.a

T;i (g)Saa (g)

=L

i,a

T;i(g) L i

Saa(g)

a

(24.44) So the character of the tensor product is the prodnct of the individual characters. An important specialcase is the tensor product of a representationwith itself. For such a representation, the matrix elements satisfy the symmetryrelation (T <8l T)ia,jb(g) = (T <8l T)ai,bj(g). This symmetry can be used to decompose the tensor product space into two snbspaces that are separately invariant under the action of the group. To do this, take the span of all the symmetric vectors of the form (Iviwj) + IVjwi» E V <8l V and denote it by (V <8l Vis. Similarly, take the span of all the antisymmetric vectors of the form (Iviw j) - IVjwi}) E V <8l V and denote it by (V <8l Via. Next note that IViWj) = !(Iviwj)

+ IVjwi» + !(Iviwj) -Ivjwi».

24.7 TENSOR PROOUCT OF REPRESENTATIONS

701

It follows that every vector of the product space can be written as the sum of a symmetric and an antisymmetric vector. Furthermore, the only vector that is both symmetric and antisymmetric is the zero vector. Therefore,

V ® V = (V ® V), Eil (V e V)a. Now consider the action of the group on each of these subspaces separately. From the relation T® T(g) IViWj) es T ® T(g)(lvi) , IWj))

=

(~Tki(g) IVk) , ~ Tlj(g) IWI))

= L

niTlj(g)(g) (IVk) ,

IWI)) =

k,/

L(T

e T)kl.ij(g) IVk W/)

k,1

weobtaio T ® T(g)(lviWj)

= L

[(T

±

IVjWi)

e T)kl,ij (g) ± (T ® T)kl,ji (g)] IVkWI). (24.45)

k,/

Kronecker product reduces to the symmetric and the antisymmetric representations

Problem 24.21 shows that the RHS can be written as a sum over the symmetric (for the plus sign) or antisymmetric (for the minus sign) vectors alone. It follows that 24.7.1. Box. The Kronecker product of a representation with itself is always reducible into two representations, the symmetrized product and the antisymmetrizedproduct representations.

24.7.1 Clebsch-Gordan Decomposition A common situation in quantum mechanics is to combine two simple systems into a composite system and see which properties of the original simple systems the composite system retains. For example, combining the angular momenta of two particles gives a new total angnlar momentum operator. The question of what single-particle angnlar momentum states are included in the states of the total angular momentum operator is the content of selection rules and is of great physical interest: A qnark and an antiquark (two fermions) with spin! always combine to form a mesou (a boson), because the resnltiug composite state has no projection onto the subspace spanned by half-integer-spin particles. In this section, we study the mathematical foundation of this situation. The tensor product of two irreducible representations T(a) and T(jJ) of G is denoted by and it is,

rv-».

702

24. GROUP REPRESENTATION THEORY

in general, a reducible representation. The characters, generally compound, are denoted by X(ax Pl. Equation (24.14), combined with Equation (24.44), tells us what irreducible representations are present in the tensor product, and therefore onto which irreducible representations the product representation has nonzero projection:

xi

a xPl

= xi

a1

r

. X'
=L

m':/

xia1,

0'=1

where m~fJ are nonnegative integers. We rewrite this more conveniently in terms of vectorsas r

Ix(ax P1} =

Lm,:/ Ix(a l}, 0'=1

map a

simply reducible group

= _1_ (x(all X(ax Pl) = _1_ ~ c'X-(alX(alX(/ll IGI

A group for which

IGI~

I

I

I

I



(24.46)

m':/ = 0, 1 is called simply reducible.

Rudolph Friedrich Alfred C1ebsch (1833-1872) studied mathematics in the shadow of Jacobi at the University of Konigsberg, two of his teachers having been students of Jacobi. After graduationhe held a number of positions in Germany, including positions at theuniversities of Berlin, Giessen, and finally Gottingen, where he remained untilhis death. He and Carl Neumann, son of one of theaforementioned Jacobian teachers, founded the Mathematische Annalen. Clebsch began his career in mathematical physics,producing a doctoral thesison hydrodynamics anda bookon elasticity in which hetreated theelasticvibrations ofrodsand plates. Theseworks were primarily mathematical, however, andhe soonturned his attention more topure mathematics. His links to Jacobi gave rise to his first work in that vein, concerning problems in variational calculus and partial differential equations, in which he surpassed theresults of Jacobi's work. Clebsch first achieved significant recognition forhis work in projective invariants and algebraicgeometry.He was intrigued by the interplay between algebra and geometry,and, sincemany results in thetheory of invariants havegeometric interpretations, thetwo fields seemed natural choices.

24.7.2. Example. Referringto Table24.8 of Problem 24.15, and nsing Eqnation (24.44), we can constructthe compoundcharacter IX(4x5)} with components 9, -1,1,0, -1. Then,

24.7 TENSOR PRODUCT OF REPRESENTATIONS

703

we have

Forthe first irreducible representation, we get 5 45 _ (1) (4X5») _ (1)* (4x5) ml - 24 X X - 24~CIXi Xi

I

-.!.- (

-.!.-" . 1=1

1 = 24[1·1 ·9 + 6·1 . (-1) + 3·1·1 + 8·1 ·0 + 6·1 . (-1)] = O. Forthe secondirreducible representation, we get

m~5 = ~ (x(2)1 X(4X5»)

= -1 [1 . 1 ·9 + 6· (-1) . (-1) + 3 . 1 . 1 + 8 . (-1) . 0 + 6 . (-1) . (-1)] = 1. 24

m1

Similarly, mj5 = 1, 5 = I, and m~5 = 1. Wethussee thattheidentity representation is notincluded in thedirect product of irreducible representations 4 and5; all other irreducible representations of 84 occuronce in T(4x5). II

Clebsch-Gordan series

In terms of representations themselves, we have the so-called ClebschGordan series r

T(ax~)(g)

=

L m~~T(U\g), cr=l

where we have used Equation (24.13) How to obtain invariants from the product of representations

24.7.3. Example. Theone-dimensional identity representation playsa special rolein the application of group theory to physics becauseany vector(function) in its carrier spaceis invariant under theactionof thegroup, andinvariant vectors oftendescribe specialstates of the quantum mechanical systems.Forexample, the ground state of an atomicsystemwith rotational invariance has zero orbital angular momentum, corresponding to a spherically symmetric state. Another example comes fromparticle physics. Quarks areusuallyplacedin the states of an irreducible representation of a group [SU(n), where n is the number of "flavors" such as up, down, charm], and antiquarks in its adjoint. A question of great importance is whatcombination of quarks andantiquarks leads to particles---called singlets-that are aninvariant of the group. Forthe case of quark-antiquark combination, the answer comes in the analysis of the tensor product of one irreducible representation, say r(a), andone adjointrepresentation, say t(fJ). In fact, using Equatiou(24.47), we have 1

r

IGI

~ l=l

1

r

IGI

~ 1=1

ma~ = _ " c'x-(l)X(a)X-(~) = - " c'X(a)X-(~) = ~ 1

I

l

I

I

l

I

Q

lap'

704

24. GROUP REPRESENTATION THEORY

whereweusedEquatiou(24.13) andthefactthat all characters of theidentityrepresentation areunity. Thus to construct aninvariant state, we need to combine a representation with its adjoint, in which case we obtain theidentity representation only once. II

Paul Albert Gordan (1837-1912), the son of David Gordan, a merchant, attended gymnasium and business school, then workedfor severalyears in banks. His early interest in mathematics was encouraged by the private tutoring he received from a professorat the Friedrich Wilhelm Gymnasium. He attendedErnst Kummer's lecturesin number theoryatthe University of Berlin in 1855, then studied at the universities of Breslau, Konigsberg, andBerlin.At Konigsberg he came underthe influenceof Karl Jacobi's school, andatBerlinhis interestin algebraicequationswas aroused. His dissertation (1862), which concernedgeodesics on spheroids, receiveda prize offeredby the philosophy facultyof the Universityof Breslau. The techniques thatGordan employed in it were those of Lagrange andJacobi. Gordan's interest in function theory led him to visit G. F. B. Riemann in GOttingen in 1862, but Riemann was ailing, and their association was brief. The following year, Gordan was invited to Giessen by Clebsch, thus beginning the fruitful collaboration most physicists recognize. Together they produced work on the theory of Abelian functions, based on Riemann's fundamental paperon thattopic, and severalof Clebsch's papers are consideredimportant steps toward establishing for Riemann's theories a finn foundation in tenus of pure algebraic geometry. Of conrse, the Clebsch-Gordan collaboration also produced the famous coefficients thatbeartheirnames, so indispensable to the theoryof angularmomentom coupling foundin ahnusteveryareaof modemphysics.In 1874 Gurdan becameaprofessoratErlangen, wherehe remained untilhis retirement in 1910. He married SophieDeuer, the danghterof a Giessenprofessorof Romanlaw,in 1869. In 1868 Clebsch introduced Gordan to thetheoryof invariants, which originated in anobservation of George Boole's in 1841 andwasfortherdeveloped by ArthnrCayley in 1846. Following the workof these twoEnglishmen, a Germanbranchof the theorywas developed by S. H. Aronhold and Clebsch, the latter elaborating the former's symbolicmethodsof characterizing algebraic forms and theirinvariants. Invariant theorywas Gordan's main interestfor the rest of his mathematical career; he becameknownas the greatestexpertin the field, developingmany techniques for representing andgenerating forms andtheirinvariants. Gordan madeimportant contributions to algebraand solutions of algebraic equations, and gave simplifiedproofs of the transcendence of e and x . The overallstyle of Gordan's mathematical work was algorithmic. He shied away from presenting his ideas in informal literaryforms. He derived his resultscompntationally, workingdirectlytowardthe desired goal withoutofferingexplanations of the concepts thatmotivated his work. Gordan's only doctoral student, Emmy Noether, was one of the firstwomen to receive a doctorate in Germany. She carried on his work in invariant theoryfor a while, butunder the stimulus of Hilbert's school atGOttingen herinterests shiftedandshe becameone of the primary contributors to modem algebra.

24.7 TENSOR PROOUCT OF REPRESENTATIONS

705

So far, we have concentrated on the reduction of the operators and carrier spaces into irreducihle components. Let us now direct our attention to the vectors themselves. Given two irreducible representations T{a) and T{P) with carrier spaces spanned by vectors {14>!a»))~:l and

(1",jP»))~:l' we form the direct product

representation T{a xP) with the carrier space spanned by vectors

Clebsch-Gordan coefficients

(14)!a) '"jP»)). We

know that T{axp) is reducible, and Equation C24.47) tells us how many times each irreducible factor occurs in T{axp). This means that the span of (14)!a),,,jP))) can be decomposed into invariant irreducible subspaces; i.e., there must exist a basis of the carrier of the product space the vectors of which belong to irreducible representations of G. More specifically, we should be able to form the linear combinations

I"'t),q) = L

CCap; a,

q lij; k) I4>!a)",jP») ,

C24.48)

ij

which transform according to the rows of the c th irreducible representation. Here the subscript k refers to the row of the crth representation, and q distinguishes among functions that have the same o and k, corresponding to the case where 2. For simply reducible groups, the label q is unnecessary. The coefficients CCap; a, qlij; k) are called the Clebsch-Gordan coefficients for G. These coefficients are normalized such that

m':! :::

L C*Cap; a, qlij; k)CCap; a', q'lij; k') = 8uu,8qq,8kk" ij

L C*Cap; a, qlij; k)CCap; a, qli' t'. k) = 8w8jj'. aqk

I"'ku),q)

This will guarantee that are orthonormal if the product vectors form an orthonormal set. Using these relations, we can write the inverse ofEquation C24.48) as

14>!a)",jP») =

L

C*Cap;

a, qlij; k)

I"'ku),q} .

C24.49)

oqk

24.7.2 Irreducible Tensor Operators An operator A acting in the carrier space of the representation of a group G is transformed into another operator, A f--* TgAT; 1, by the action of the group. Just as in the case of vector spaces, one can thus construct a set of operators that transform among themselves by such action and lump these operators in irreducible sets. irreducible set of operators

24.7.4. Definition. An operator A!a) is said to to transform according to the ith row of the ath irreducible representation if there exist n a - lather operators

706

24. GROUP REPRESENTATION THEORY

(24.50)

where (T~'\g)) is the matrix representation of g. The set of such operators is called anJirreducible set of operators (or irreducible tensorial set). scalar operator

In particular, if T,J'\g) = 8ij, i.e., if the representation is the identity repre-

sentation, then A= TgAT;I, and Ais called a scalaroperator. The term "scalar" refers to the fact that A has ouly one "component," in contrast to the other operators of Equation (24.50), which may possess several components. Consider the set of vectors (functious) defined by IVrij} == Ai') lj~)), where

I(/l)) transform according to the ,Bth irreducible represeutation. These vectors J transform accor ding to nu

np

k~1

1~1

Tg IVrij) = TgAi')T;I Tg 1<1>1)) = LJk\'\g}Ak·) L =

Tz1\g} I?))

L Tk\')(g)Tz1)(g)Ak') I?)) = LTk\:rf!(g) IVrkl) , U

~.5D

U

i.e., according to the representation T('X~). This means that the vectors IVrij) have the same transformation properties as the tensor product vectors Therefore, usiug Equation (24.49), we can write

Ai') lj~)) =

li') Vrf)).

L C*(a,B; rr, qlij; k) Iw~D'),q), aqk

and more importantly,

(<1>,\1')1 Ai') I(~)) = J

L C*(a,B; a, qlij; k) (<1>,\1')1 wt),q) oqk

=

L C*(a,B; y, qlij; m) (<1>,\1') q

Wigner-Eckart theorem and reduced matrix elements

" ' use Eq. (24.38) here 1

w,\1'),q). (24.52)

It follows that the matrix element of the operator Ai') will vanish uuless tJie irreducible representation T(Y) occurs in the reduction of the teusor product T(') ® T(~), and this can be decided from the character tables and the ClebschGordan series, Equation (24.47). There is another remarkable property of Equation (24,52) that has significant physical consequences. Notice how the dependence ou i and j is contained entirely in the Clebsch-Gordan coefficients. From Equation (24.39) it follows that

24.8 REPRESENTATIONS OF THE SYMMETRIC GROUP

707

I

(q,;,;) >It;';).q) is independent of m. Therefore, this dependence must also be contained entirely in Clebsch-Gordan coefficients. One therefore writes (24.52) as

reduced matrix element

(24.53)

This equation is known as the Wigner-Eckart theorem, and the numbers multiplying the Clebsch-Gordan coefficients are known as the reduced matrix elements. From the point of view of physics, Equation (24.53) can be very useful in calculating matrix elements (expectation values and transition between states), once we know the transformation properties ofthe physical operator. For example, for a scalar operator S, which, by definition, transforms according to the identity representation, (24.53) becomes

i.e., scalar operators have no matrix elements between different irreducible representations of a group, and within an irreducible representation, they are multiples of the identity matrix. This result is also a consequence of Schur's lemma.

24.8 Representations of the Symmetric Group The symmetric (permutation) group is an important prototype of finite groups. In fact, Cayley's theorem (see [Rotm 84, p. 46] for a proof) states that any finite group of order n is isomorphic to a subgroup of Sn. Moreover, the representation of Sn leads directly to the representation of many of the Lie groups encountered in physical applications. It is, therefore worthwhile to devote some time to the analysis of the representations of Sn.

24.8.1 Analytic Construction The starting point of the construction of representations of the symmetric group is Equation (24.35), which is valid for any finite group. There is one simple character that every group has, namely, the character of the one-dimensional symmetric representation in which all elements of the group are mapped to 1 E R. Setting ~y) = I in (24.35), and noting that Lj dj = ds, we obtain

1{rH = IGldi

, - IHlci'

(24.54)

where {1{riH} are the components of a compound character of G. Frobenius has shown that by a clever choice of H, one can completely solve the problem of the construction of the irreducible representations of Sn. He proceeded

708

24. GROUP REPRESENTATION THEORY

as follows. Consider a partition (A) = (AI•...• An) ofn. The symmetric groups IS).,} are subgroups of Sn and have no elements in common-therefore they all commute with one another. The direct product of these subgroups is a subgroup of Sn. which we denote by S().):

S().) = S)., x···

X

S).".

If we denote the compound character of Equation (24.54) by 11/J().)) in this case. calculate IHI. Ci. and di. and substitute the results in (24.54}--all of which can be done in closed form-we obtain an explicit formula for the components of 11/J().)). Tbis formula is messy. and we shall not derive it here. The interested reader may refer to [Harne 89. pp. 189-192] for details. We are really interested in the simple characters of Sn. and Frobenius carne up with a powerful method of calculating them. Since there is a one-to-one correspondence between the irreducible representations and conjugacy classes. and another one between conjugacy classes of Sn and partitions of n. we shall label the simple characters of Sn by partitions of n. Thus. instead of our common notation a ). we use x&~). where (A) denotes a partition of n, and (I) a cycle structure of

xi

s;

Suppose we want to find the irreducible characters corresponding to the cycle structure (I) = (I a • 2P• 3Y • • • • ). These form a column under the class (I) in a character table. To calculate the irreducible characters. form two polynomials in (XI. X2• . . . • xn) as follows. The first one. which is completely symmetric in all

variables, is

8(1). == (n?=Xi )a (n?=Xf )P (n?=xi) Y l=l

1=1

(24.55)

1=1

The second one is completely antisymmetric, and can be written as

I D(Xl ..... x n ) sa det

I

I =IT(Xi-Xj) i
(24.56) It can be shown that the simple characters of Sn are coefficients of certain terms of the product of these polynomials. To be exact. we have

(24.57)

24.8 REPRESENTATIONS OF THE SYMMETRIC GROUP

709

The outer sum goes over all partitions of n, the inner sum over all permutations of Sn. The procedure for finding the simple characters of Sn should now be clear from (24.57):

24.8.1. Box. To find the simple character xg~iPA~\, construct the corresponding symmetric and antisymmetricpolynomials. multiply them together, collect all terms ofthe form Al+n-l A2+n-2 An_l+l An X,,(2) ••• X,,(n_l) x,,(n)

X,,(I)

for all permutations 11: E Sn; the coefficient ofsuch a term will be the desired character. n-2 · th . f A1+ n- 1 x A2+ An_1 +1 An < . 'di"" Smce e coeffi cients 0 x1r (l ) ... xrr(n_l) x 1r(n) ror VarIOUS n suer rr(2) by a sign, in practice, the coefficient of just x~l+n-l ... X;~11+1x;n is precisely

the simple character we are looking for.

24.8.2. Example. The best way to understaod the procedure described above is to go through an example in detail. Wecalculate the characters of 83 using the above method. Label the rows of the character table with the partitions of 3. These are (3), (2,1), and (1,1, I). Similarly, label the colwnns with the conjugacy classes, or cycle structures: (1 3) , (1,2), and (3). The first cycle structure has a 3, {3 0 y. Therefore,

=

= =

_ (

S(l3) -

)3 _ 3 3 3 XI +X2+ x3 -XI +x2 +x3

+ 3(xrX2

+ XrX3 + XIX~ + X~X3 + XIX~ + X2X~) + 6xj X2X3 (24.58)

and

so that

(24.59)

=

=

=

=

Now we oote that for (A) (3), AI 3, A2 0, and A3 O. Therefore, the n - 1 A2+n-2 . f A1+ An 5 . (3) Simi!' I • coefficiento Xl X "'X =X1X2gIVeSX(l3)' any. ror (') A = (2 , 1.0). n

2

Al = 2, ).,2 = 1, and).3

X&i;).

=

0, and the coefficient of x~1+n-lx~2+n-2... x;n

=

=

=

= xtxi

=

gives Finally, for (A) (1, I, I), Al A2 A3 1, and the coefficient of n-1 n-2 A1+ A2+ An 3 2 . (1,1.1) Th . Xl X2 ... Xn = Xl X2X3 grves X(!3) . ese coefficients can be read 0 ff by scanning through Equation (24.58) while multiplying its terms by those ofEquation (24.59) and keeping track of the coefficients of the products of therelevant powers of xj , x2. and x3. Thereader mayverify that there is only oneterm of theform. x~ X2. whosecoefficient is

710

24. GROUP REPRESENTATION THEORY

(1 3 )

(1,2)

(3)

T(3)

1

1

1

T(2.1) T(I,I,I)

2

0

-1

1

-1

1

Table 24.5 The character table for 83. Each column corresponds to a coojugacy class, eachrowto a partition of 3. Thelasttwo rows havebeen switched compared to Table 24.4.

1, giving

xgJ) =

1; there aretwo terms of the form

xixi. whose coefficients are -1 and

3, giving Xg3~) = 2; and there arefour terms of the form XiX~X3, whosecoefficients are +1, -3, -3, and +6. giving xg;~,I) = 1. Therefore, the first column of the character table

of 83 is with a

(i). Toobtain the second column, we consider the second conjugacyclass, 0,2), = I = fJ and y = O. The corresponding symmetric polynomial is

8(1,2) = (XI + x2 + X3)(X[ + x~ + xl) 333222222 = Xl +x2 +x 3 +X 1X2 +xIX3 +Xlx2 +x2x3 +Xlx3 +X2 x3'

(24.60)

V(XI, x2, X3) is the sarne as before. Multiplying and keeping track of the coefficients of 5 4 2 3 2 . (3) (2,1) (1,1.1) XI x2, XI x 2' and Xl x 2x3, we obtam X(l,2) = I, X(l,2) = 0, and X(l,2) = -1. It follows

that the secondcolumnof the character tableof 83 is

(~1)' = 0 = fJ, and y = 1. Therefore,

The last column is obtained similarly. We note that a the synnnetric polynomial is 8(3)

33 = XI3 +x2 +x3'

and theantisymmetric polynomial is thesameasbefore. Multiplying thesetwopolynomials & (3) (2,1) (1,1,1) . th and extracting e coeffici Cleot s as b erore, we get X(3) = I ,X(3) = - I ,and X(3) =1. It follows that the third column of the character table of 83 is (

~I ).

Collecting all thedata obtained above,we canreconstruct thecharacter tableof 83. This is shown in Table 24.5. The irreducible representations are labeledby the three possible partitions of 3, andtheconjugacy classes by the three cycle structures. II

24.8.2 Graphical Construction The analytic construction ofthe previous subsection can be handled using graphical techniques that are considerably simpler. To begin with, let us find the character of the identity element of Sn. The cycle structure is (In), i.e., all cycles consist of a single element. Thus, " = n, and {J, y, etc. are all zero. It follows that the LHS of Equation (24.57) is (Lxi)n D(xj). We calculate this product one power of

2>i

24.8 REPRESENTATIONS OF THE SYMMETRIC GROUP

at a time. For the same reason as in the example above, of

711

xg2) will be the coefficient

Ferdinand Georg FrobeDins (1849-1917), the son of a parson, was born in Berlin andbeganhis mathematical studies at G5tti.ngenin 1867.Hereceivedhis doctorate inBerlinthree years later. Four years later, on thebasisof his mathematical research, he was appointed assistant professor attheUniversity of Berlin. He achieved the rank of full professor at the Eidgeniissische Polytechnikum Ziirich beforereturning to Berlin as a professor of mathematics in 1892. During the earlyyears of Frobenius's career, modem group theory was in its infancy. He combined its three main branches of study-the theory of solutions to algebraic equations (permutation groups and the workof Galois), geometry (transformation and Lie groups), and number theory-to produce the concept of the abstract group. He collaborated with Issai Schur in representation andcharacter theory of groups. His paper Uber die Gruppencharactere is of fundamental importance. It waspresented to the Berlin Academy on 16July 1896 and it contaius work that Frobeuius had douein the preceding few months. In a seriesof letters to Dedekind, thefirst on 12 April 1896. his ideas on group characters quickly develop, andFrobenius is able to construct a complete set of representationsby complex uumbers. In a letter to Dedekind on 26 Apri11896 Frobeuius finds the irreducible charactersfor the alternating group, and the symmetric groups. Iu 1897Frobeuiusreformulatedthe work of Molien-the Latvian studeutof Klein,who, in histhesis, classified thesemisimple algebras usingthemethod of group rings-in terms of matrices and thenshowed that his characters arethetraces of the irreduciblerepresentations. Frobenius's character theory found important applications in quantum mechanics and was used with great effect by Burnside, who wrote it up iu the 1911 edition of his Theory of Groups of FiniteOrder. Frobenius is alsoremembered as the originator of a seriesmethod for solvingordinary differential equations. Despitethe clearlygreater importance of his workin group theory, this method of Frobenius serves admirably to perpetuate his name.

If we multiply D(xj) by 2:Xi one X at a time, we increase the power of one of the Xi'S by one. If at any stage, two of the exponents become equal, the term must vanish, due to the antisymmetry of (2: Xi )D(x j)' Therefore, as we raise the degree of the polynomial by one at each stage, the power of Xl must be raised at least as fast as X2, and the power of X2 must be raised at least as fast as X3, etc. Our goal is to raise the power of Xl by AI, that of X2 by A2, and, in general, the power of Xi by Ai, making sure that at each stage, the number of multiplications by Xl is greater than or equal to the number of multiplications by X2, etc. The total number

712

24. GROUP REPRESENTATION THEORY

xg2),

of ways by which we can reach this goal will be which is also the dimension of the irreducible representation (A) by Equation (24.9). To see the argumentmore clearly, suppose that we are interested in the dimension of the irreducible representation of S4corresponding to (3, I). Then we must raise the power of Xl by 3 and the power of X2 by I; X3 and X4 will remain intact, and therefore will not enter in the following discussion. It follows that D(xj) is to be multiplied by xi X2, one x-factor at a time, the number of Xl-factors always exceeding the number of x2-factors. The possible ways of doing this are xix2'

XrX2XI,

(24.61)

XIX2Xr.

Note that as we count the factors from left to right, the number of XI'S is always greater than or equal to the number of X2 'so Thus x2xi is absent because X2 occurs without Xl occurring first. It follows that the dimension of the irreducible representation (3, I) is 3. A graphical way to arrive at the same result is to draw AI = 3 boxes on top and A2 = I box below it:

DOD

o

The next step is to fill in the boxes with numbers corresponding to the position of Xl (filling up the first row) and X2 factors (filling up the second row) in Equation (24.61). Since in the first term of (24.61), the XI'S occupy the first, second, and third positions, we enter I, 2, and 3 in the first row, and 4 in the second row corresponding to the last position occupied by X2. Similarly,in the second term of (24.61), the XI'S occupy the first, second, and fourth positions; therefore, we enter I, 2, and 4 in the first row, and 3 in the second row corresponding to the position occupied by X2. Finally, in the last term of (24.61), the XI'S occupy the first, third, and fourth positions; therefore, we enter I, 3, and 4 in the first row, and 2 in the second row corresponding to the position occupied by X2. The result is the graph shown below:

Young frame defined

24.8.3. Definition. Let (A) = (AI, A2, ... , An) be a partition of n. The Young frame (or the Young pattern) associated with (A) is a collection ofrows ofboxes (squares) aligned at the left such that the first row has AI boxes, the second row A2 boxes, etc. Since Ai 2: AHJ, the length ofthe rows decreases as one goes to the bottom ofthe frame.

xt' ...x;",

The Youngframe associated with (A) represents which multiplies the antisymmetric polynomial D(xj). To find the dimension of the irreducible representation T(A), we have to count the number of ways in which the x-factors can be permuted among themselves such that as we scan the product, the number of Xi'S is never less than number of X /s if j > i. This leads to

24.8 REPRESENTATIONS OF THE SYMMETRIC GROUP

standard Young tableaux defined

713

24.8.4. Definition. A standard Young tableau (or diagram, or graph) is a Young frame filled with numbers 1 through n such that 1. the numbers are placed consecutively left to right on the rows starting with 1 in the far-left box ofthe first row; 2.· no box ofany row is to be filled unless all boxes to its left are already filled; 3. at each stage, the number of boxes filled in any row is never less than the number ofrows below it.

regular graphs

Tableaux satisfying the last condition are called regular graphs.

It followsthat in a Young tableau,the number 1is alwaysin the upperleft-hand box, and that going down in a column, the numbers must increase.

24.8.5. Theorem. Let (A) be a partition ofn. Then the dimension ofthe irreducible representation T().) is equal to the number ofstandard Young tableaux associated with (A). 24.8.6. Example. We wish to calculate the dimension of each irreducible representation of S4.The partitions are (4), (3, I), (2, 2), (2, I, I), and (1,1, I, 1) whose associatedYoung framesare shownbelow:

ODDO

(4)

DDD

DD DD

(3,1)

(2,2)

o

DD

o o o o

(2,1,1)

(1,1,1,1)

D D

The number of standard Young tableaux associated with (A) = (4) is I, because there is only one way to place the numbers 1 through 4 in the four boxes. Thus, the dimension of T(4) is 1. For (A) = (3, I), we can place 2 either to the right of I or below it. The first choice gives rise to two possibilities for the placement of 3: Either to the right of 2 or below 1. The second choice gives rise to only one possibility for 3, namely to the right of 1. With 1,2, and 3 in place, the position of 4 is predetermined. Thus, we have 3 possibilities for (A) = (3, i), and the dimension of T(3,1) is 3. For (A) = (2,2), we can place 2 either to the right of I or below it. Both choices give rise tu ouly one possibility for 3: In the first case, 3 can ouly go noder I; in the second case to its right. With I, 2, and 3 in place, the position of 4 is again predetermined. Thus, we have 2 possibilities for (A) = (2,2), and the dimension of T(3, 1) is 2. The reader may check that the dimension of T(2, I, 1) is 3, and that of T(I, I, I,I) is 1. Figure 24.1 summarizes these findings. We note that the dimensions satisfy 12 + 32 + 22 + 32 + 12 = 24, the second equation of (24.18). l1li

714

24. GROUP REPRESENTATION THEORY

(4)

(3,1)

(2,2)

(2,1,1)

[IJmwl}] [TI[1][] I}]

[IJ[]I}]

[TIm

[IJ[]

[TIm

[TIl}] [1] DJ

[]rn [] [±]

n(3,1) =

3

rn

rnrn

n(2,1,1) = 3

n(l,I,I,I) = I (1,1,1,1)

Figure 24.1 The standard Young tableaux, andthe dimensions of irreducible representationsof 84.

24.8.3 Graphical Construction of Characters The product of the symmetric polyuomial S(I) and the antisymmetric polynomial D(xj) contains all the information regarding the representations of S«. We can extract the simple characters by looking at the coefficients of appropriate products of the x-factors. This can also be done graphically. Without going into the combinatorics ofthe derivation of the results, we state the rules for calculating the simple characters, and exantioe one particular case in detail to elucidate the procedure, whose statement can be very confusing. As before, we label the irreducible representations with the partitions of n. However, we separate out the common factors in a cyclic structure, labeling the cycles by li, 12, etc. For example, (2, 12) has li = 2, 12 = I, and 13 = I. So, we write (I) as (li, 12, ... , 1m ) .

regular application

24.8.7. Definition. A regular opplkation ofr identical symbols to a Youngframe is the placement of those symbols in the boxes of the frame as follows. Add the symbols to any given row, starting with the the first (farthest to the left) unoccupied cell, until the symbols are all used 6 or the number exceeds that of the preceding

6n is understood thatif you cannot placeall symbols on thefirst row, thenyou should start atthesecondrow.

715

24.8 REPRESENTATIONS OF THE SYMMETRIC GROUP

positive and negative applications

line by one. In the latter case, go to the preceding line and repeat the procedure, making sure that the final result of adding all r symbols will be a regular graph. If in this process the r symbols are divided among an even number of rows, we speak of negative application. If the number of rows is odd, we have a positive application. As an illustration, considerthe regular application offive 2's to the blank Young frame shown below.

0000 000

o o

We cannot start on the first row because it does not bave enongh boxes for the five 2's. We can start on the second row and put one 2 in the first box. This brings the number of 2's in the second row to one more than in the first row; therefore, we should now go to the first row and put the rest of the symbols there. We could start at the third row, put one 2 in the first box, put a second 2 in the first box of the second row, and the rest in the first row. Altogether we will have 3 regular applications of the five 2's. These are shown in the diagram below.

rnrnrno rnrnoo moo moo rn m o rn Of these the first and the last tableaux are negative applications, and the middle

one is positive.

24.8.8.Theorem. The character ofthe irreducible representation T(A) ofthe class (I) = (lJ,12, ... ,1m) is obtained by successive regular applications oflJ identical symbols (usually taken to be 1's), then lo, identical symbols of a different kind (usually taken to be 2's), etc. The character is then equal to the number of ways of building positive applications minus the number of ways of building negative applications.

x(W

The order in which the Ij's are applied is irrelevant. However, it is usually convenient to start with the largest cycle. The best way to understand the procedure is to construct a character table. Let us do this for 54. As usual, the rows are labeled by the various partitions (A) of4. We (22) , (2, 1, I) (2, 12) , (1, I, I, 1) (14) . choose the order (4), (3, I), (2,2) The columns are labeled by classes (I) in the following order: (14 ) , (2, 12 ) , (22 ) , (3, I), (4), where, for example, (2,1 2) means that 11 2,12 I, and 13 I. Example 24.8.6 gives us the first column of the character table. Similarly, the first row has 1 in all places, because it is the trivial representation. Our task is therefore

=

Oetailed analysis of the construction of the character teble for 5,

=

=

=

=

=

716

24. GROUP REPRESENTATION THEORY

to fill in the rest of the table one row at a time. The second row, with (A) = (3, 1), has a Young frame that looks like

DOD

o

and for each class (column) labeled (11, ... ,1m ) , we need to fill this in with 11 identical symbols (1's), 12 identical symbols of a different kind (2's), etc. The second column has 11 = 2, lz = 1 = 13. So we have two l.'s, one 2, and one 3. Ifwe start with the first row, the two I's can be placed in its first two boxes. If we start with the second row, the two l 's must be placed vertically on top of each other. In the first case, we have two choices for the 2: Either on the first row next to the two 1's, or on the second line. In the second case, we have only one choice for the 2: in the first row next to 1. With 1's and 2 in place, the position of 3 is determined. The three possibilities are shown below:

The first two are positive applications, the third is negative because the 1's occupy an even number of rows. We therefore have (3,1) _

X(2,12) -

+1 + 1 - 1 -- +1.

The third colunm has 1J = 2 = 12. So we have two 1's and two 2's. We place the 1's as before. When the two 1's are placed vertically, we can put the 2's on the first row and we are done. When the 1's are initially placed in the first row, we have no way of placing the 2's by regnlar application. We cannot start on the first row because there is only one spot available (remember, we cannot go down once we start at a row). We cannot start on the second row because once we place the first 2, we are blocked, and the number of symbols in the second row does not exceed that of the first row by one. So, there is only one possibility:

Not allowed

Allowed

The only allowed diagram is obtained by a negative application of 1'So Therefore, (3,1) X(22)

=-1.

The fourth colunm has 1J = 3 and h = 1. So we have three 1's and one 2. There are two ways to place the 1's: all on the first row, or starting on the second row and working our way up until all boxes are filled except the last box of the first row. The placement of 2 will be then predetermined. The result is the two diagrams shown below:

24.8 REPRESENTATIONS OF THE SYMMETRIC GROUP

717

rnrnrn rn

rnrnrn rn

The first diagram is obtained by a positive application of I's, the second by a negative application. Therefore, (3.1)

X(3.1)

= +1 -I = O.

Finally, for the last column, It = 4. There is only one way to put all the I's in the frame, and that is a negative application. Thus, X~i1) = -l. Rather than going tbruugh the rest ofthe table in die same gory detail, we shall point ont some of the trickier calcnlations, and leave the rest of the table for the reader to fill in. One confusion may arise in the calcnlation of looks like this,

xg:;. The frame

00

DO

and we need to fill this with two I's and two 2's. The I's can go into the first row or the first column. The 2's then can be placed in the second row or the second column. The resnlt is

The first diagram has no negative application. The second has two negative applications, one for the I's, and one for the 2' s. Therefore, the overall sign for the second diagram is positive. It follows that

xg:; = + I + I = +2.

The calcnlation of x~)') may also be confusing. We need to place four I's in the frame. If we star! on the first row, we are stuck, because there is room for only two I's. If we star! in the second row, then we can only go up: Putting the first I in the second row causes that row to have one extra I in comparison with the preceding row. However, once we go up, we have room for only two I's (we cannot go back down). So, there is no way we can place the four I's in the (2 2 ) frame, (2') X(4)

= O. The character table for 84 is shown in Table 24.6 (see Problem 24.15 as well). The reader is urged to verify all entries not calcnlated above. The character table for 85 can also be calcnlated with only minor tedium. We quote the resnlt here in Table 24.7 and let the reader check the entries of the table. and

24.8.4 Young Operators The group algebra techniques of Section 24.4-which we used in our discussion of representation theory in a very limited way-provide a powerful and elegant tool

718

24. GROUP REPRESENTATION THEORY

T(4) T(3.1) T(2') T(2,!') T(l4)

(14)

(2,1 2 )

(22 )

(3, I)

(4)

I 3 2 3 I

I I

I -I

I -I

0 -I -I

2 -I I

I 0 -I

0 I

0 I -I

Table24.6 Cbaractertable for 54. The rows and columns are labeled by the partitionsof 4 andcycle structures, respectively.

for unraveling the representations of finite gronps. These techniqnes have been particnlarly useful in the analysis of the representations of the synnnetric group. Our emphasis on the synnnetric group is not merely due to the importance of S« as a paradigm of all finite groups. It has also to do with the unexpected usefulness of the representations of S« in stndying the representations of GL(V), the paradigm of all (classical) continuous groups. We shall come back to this observation later when we discuss representations of Lie groups in Chapter 27. To begin with, consider the element of the Sn group algebra as defined in Equation (24.22). Since mnltiplying P (on the left) by a group element does not change P, the ideal generated by P is not only one-dimensional, but all elements of Sn are represented by the number I. Therefore, the ideal AP corresponds to the (irreducible) identity representation. For Sn, there is another group algebra element that has similar properties. This is n!

Q=

LE

(24.62)

n i 11"j ,

i=l

The reader may check that and

Q2 =n!Q.

As in the case of P, Q generates a one-dimensional ideal, but a left mnltiplication may introduce a minus sign (when the permutation is odd). Thus, the ideal generated by Q must correspond to the antisynnnetric (or alternating) representation. All the irreducible representations, including the special one-dimensional cases above, can be obtained using this group-algebraic method. We shall not give the proofs here, and we refer the reader to the classic book [Boer 63, pp. 102-125]. The starting point is the Young frame corresponding to the partition (A) = (AI, ... , Am). One puts the numbers I through n in the frame in any order, consistent with tableau construction, so that the end product is a Young tableau. Let p be any permutation of a Young tableau that does not change the row

24.8 REPRESENTATIONS OF THE SYMMETRIC GROUP

T(5) TC4.1) TC3.2) TC3.1') T(2'.1) T(2.1') T(l')

(15)

(2, 13)

(22 , I)

(3,2)

(3, 12 )

(4, I)

(5)

I 4 5 6 5 4 I

I 2 I 0 -I -2 -I

I 0 I -2 I 0 I

I -I I 0 -I I -I

I I -I 0 -I I I

I 0 -I 0 I 0 -I

I -I 0 I 0 -I I

719

Table 24.7 Character tablefor S5. The rows and columnsare labeledby the partitionsof 5 andcycle structures, respectively.

horlzonlal and vertical permutations

of the numbers it acts on. Such a p is called a horizontal permutation. Similarly, let q be a vertical permutation of the Young tableau, i.e., a permutation that does not change the column of the nnmbers it acts on. 24.8.9. Definition. Consider the kth Young tableau corresponding to the partition A (A). Let the Young symmetrizer PF') and antisymmetrizer ) be the elements ofthe group algebra of Sn defined as

Qk

Young operators defined

PkCA) = ""' L..P' p

Then, the Young operator Y1A) ofthis tableau, another element ofthe group algebra, is given by Y1A) = Qf) PF').

It can be shown that the following holds. 24.8.10. Theorem. The Young operator YF') is essentially idempotent, and generates a minimal left ideal, hence an irreducible representation for Sn. Representations thus obtained from different frames are inequivalent. Different tableaux with the same frame give equivalent irreducible representations. In practice, one usually chooses the standard Yonng tableaux and applies the foregoing procedure to them to obtain the entire collection of inequivalent irreducible representations of Sn. We have already seen how to calculate characters of Sn employing both analytical and graphical methods. Theorem 24.8.10 gives yet another approach to analyzing representations of Sn. For low values of n this technique may actually be used to determine the characters, but as n grows, it becomes unwieldy, and the graphical method becomes more manageable.

720

24. GROUP REPRESENTATION THEORY

24.8.11.Example. Let us apply this method to S3. The partitions are (3), (2, I), and (13). There is only one standard Young tableau associated with (3) and (13). Thus,

wherewe havedivided theseYoung operators by 6 to makethemidempotent; we havealso used the notatioo of Example 23.4.1. One can show directly that y(3) y(l') = O. In fact, one can prove this for general Sn (see Problem 24.30).

Forthepartition (2, 1), there aretwo Young tableaux. Thefirst one has the numbers 1 and 2 in the first row and 3 in the second. In the secondtableau the numbers 2 and 3 are switched. Therefore, usingthemultiplication table for 83 as givenin Example 23.4.1, we have y?,I) = Q\2,1) p?,l) = (e - "3)(e y?,I) =

+ "2) = e + "2 -

"3 - "6,

Q~2,1) p?,I) = (e -'''2)(e + "3) = e - "2 + "3 - "5.

The reader may verifythat the product of anytwo Young operators corresponding to different Young tableaux is zeroandthat (2,I )y (2,1) _ 3y(2,1) Y1 1 1 •

Let us calculate the left ideal generated by these four Young operators. We already know from our discussion at the beginning of this subsection that £,(3) and £,(I'), the ideals generated by y(3) and y(l') , are one-dimensional. Let us find £,\2, I) , the ideal generated by y?' I) This is the span of all vectors obtained by multiplying y?' I) on the left by elemeuts

of the group algebra. It is sufficient to multiply y?' I) by the basis of the algebra, uamely

thegroup elements: (2,1) _ y(2,1)

e Y1

-

1

'

?'I) = "2 + e -

"2 Y

"3 y ? ,I )

"5 - "4 ea X\2, I) ,

= "3 +"6 " 4 YI(2, I) = "4 + "5 "5 y?,1) "6 y?,I)

= "5 + "4 = "6 + "3 -

e-

"2 = -y?,I),

"6 - "3

"2 - e

= - X\2, I) + y?, I),

= _X\2,1),

"4 - "5

= X\2,1) _

y?,I)

{Y?'

It follows from the above calculation that £,~2, I). as a vector space, is spanned by I), 2 I)}, andsince thesetwo vectors arelinearly independent, L ,1) is a two-dimensional

x?'

i

minimal ideal corresponding to a two-dimensional irreducible representation of 83. One canuse thisbasis to find representation matrices andthe simplecharacters of 83.

24.8 REPRESENTATIONS OF THE SYMMETRIC GROUP

721

Theothertwo-dimensional irreduciblerepresentation of S3,equivalent to theoneabove, is obtained by constructing the ideal.ci2,1) generated by

yi

2

,1) . This construction is left

for the reader, whois also askedto verifyits dimensionality. The resolution of the identity is easily verified:

The ej's are idempotents that satisfy ejej = 0 for i

:F j.

24.8.5 Products of Representations of S« In the quantum theory of systems of many identical particles, the wave function

musthavea particularsymmetryunderexchangeof the particles:If the particlesare allfermions(bosons), the overallwavefunction must be completelyantisymmetric (symmetric). Since the space of functionsof severalvariablescan providea carrier space for the representation of any group, we can, in the case of Sn, think of the antisymmetric (symmetric) functions as basis functions for the one-dimensional irreducibleidentity(alternating) representation. Toobtainthesebasisfunctions, we applythe Young operator y(l") (or yIn») to the argrunents of any givenfunctionof n variablesto obtainthe completelyantisymmetric(or symmetric)wavefunction." Under certain conditions, we may require mixed symmetries. For instance, in the presenceof spin,the product of the total spin wavefunction and the total space wave function must be completely antisymmetric for Fermions. Thus, the space part (or the spinpart) of the wavefunctionswill, in general,have mixed symmetry. Such a mixed symmetrycorresponds to some other Young operator,and the wave functionis obtainedby applyingthat Young operatorto the argnments of the wave function. Now supposethat we have two separatesystems consistingof nl and n2 particles,respectively, whichare all assumedto be identical.As long as the two systems are not interacting, each will consist of states that are classified accordingto the irreduciblerepresentations of its symmetricgroup. When the two systemsinteract, we shouldclassifythe states of the total system accordingto the irreduciblerepresentations of all n I +nz particles.We have already encounteredthe mathematical procedure for such classification: It is the Clebsch-Gordan decompositionof the directproductof the statesof the two systems.Since the initial statescorrespondto Young tableaux, and since we are interested in the inequivalentirreduciblerepresentations, we need to examine the decompositionof the direct product of Young frames into a sum of Young frames. We first state (without proof) the procedure for such a decomposition, and then give an example to illustrate it. 24.8.12. Theorem. To find the componentsofYoung frames in the product oftwo Young frames, drawone oftheframes. In the otherframe, assign the samesymbol, 7We must make the additional assumption that the permuted functions are all independent.

722

24. GROUP REPRESENTATION THEORY

say I, to all boxes in the first row, the same symbol 2 to all boxes in the second row, etc. Now attach the first row to the first frame, and enlarge in all possible ways subject to the restriction that no two I 's appear in the same column, and that the resultant graph be regular. Repeat with the 2 's, etc., making sure in each step that as we read from right to left and top to bottom, no symbol is counted fewer times than the symbol that came after it. The product is the sum ofall diagrams so obtained. To illustrate the procedure, cousider the product

00

o

We have put two I's in the first row and one 2 in the second row of the frame to the right. Now apply the first row to the frame on the left. The result is

oOITJm oom om o

oom 00 o om ITJ rr

Now we apply the 2 to each of these graphs separately. We cannot put a 2 to the right of the I 's, because in that case, as we count from right to left, we would start with a 2 without having counted any I's. The allowed graphs obtained from the first diagram ate

oomm om

oomm o

m

Applying the 2 to the second graph, we obtain

oom oITJm

OOITJ om

m

and to the third graph gives

oom om

m

oom o

m rn

Finally the last graph yields

00

om

mrn

00

om OJ

m

24.9 PROBLEMS

723

The entire process described above is written in terms of frames as

DO D

Q9

DODD +OODO + DOD DO -DO 0 D DOD 0 DO 0 D + DOD + DO + 2 0DO 0 OD + DD 0 OD D 0 0 D

Some simple products, some of which will be used later, are given in Figure 24.2.

24.9

Problems

24.1. Show that the action of a group G on the space of functions 1/1 given by Tg1/1(x) = 1/1(g-1 . x) is a representation of G. 24.2. Complete Example 24.1.6. 24.3. Let the vector space carrying the representation of 83 be the space of functions. Choose 1/11 (x, y, z) sa xy and find the matrix representation of 83 in the minimal invariant subspace containing 1/11. Hint: See Example 24.1.8. 24.4. Let the vector space carrying the representation of 83 be the space of functions. Choose (a) 1/11 (x, y, z) == x and (b) 1/1t(x, y, z) == x 2 , and in each case, find the matrix representation of 83 in the minimal invariant subspace containing 1/11. 24.S. Show that the represeutations T, irreducible.

T, and T*

are either all reducible or all

24.6. Show that the hermitian conjugate of Equation (24.5) gives TgAt = AtT'g' Using this relation show that S sa AtAcommutes with all Tg's. This result is used to prove Schur's lemmas in infinite dimensions. 24.7. Show that elements of a group belonging to the same conjugacy class have the same characters. 24.8. Show that the regular representation is indeed a representation, i.e., that R: G -+ GL(m, IR) is a homomorphism. 24.9. Show that a left ideal is minimal if and only if it is generated by a primitive idempotent. 24.10. Let G be a finite group. Define the element P = LXEG x of the group algebra and show that the left ideal generated by P is one-dimensional.

724

24. GROUP REPRESENTATION THEORY

000 0000 000 DO 000

B

00 0 o 0

DD+B D 000+ B DODD + DOD + DO DO 0 DOD + DO 0 0 0 0 DO 0 + 0 + DO DO 0 0 0

Figure 24.2 Someproducts of Young frames for small values ofn.

Ti·)

24.11. Show that defined in Equation (24.25) commutes with all operators T~·). Hint: Consider T~·)Ti·) (T~·»)-l. 24.12. Let Ki' denote the set of inverses of a conjugacy class K, with Ci elements. (a) Show that K i , is also a class with Ci elements. (b) Show that identity occurs exactly Ci times in the product KiKi', and none in the product KiK j if j oF if [see Equation (24.23)]. (c) Conclude that

c·· lJl -

o {Ci

if j oF if, "f..f 1 J=Z.

24.13. Show that the coefficients IGldj IIHlCi of Equation (24.35) are integers. 24.14. Show that the symmetric and the antisymmetric representations of 8n are inequivalent. 24.15. Construct the character table for 84 from that of 83 (given as Table 24.4), and verify that it is given by Table 24.8. 24.16. Show that all functions transforming according to a given row of an itreducible representation have the same norm. 24.17. Show that if the group of symmetry of V contains that of Ho and ItI>;·») belongs to the jth column of the ",th irreducible representation, then so does V ItI>;·»). Conclude that (tI>;.) IV ItI>;·») = 0 for i oF j. 24.18. Find the irreducible components of the representation of Example 24.1.6.

24.9 PROBLEMS

I

T(l) T(Z) T(3) T(4) T(S)

Table 24.8

Kl

6Kz

3K3

8K4

6Ks

1 1 2 3 3

1 -1 0 1 -1

1 1 2 -1

1 1

-1

-1

-1

0 0

725

1 0 -1 1

Character table for84.

24.19. Showthat p(3) 11fr3} of Example 24.6.3 is a linear combinationof p(3) 11fr1) and p(3) l1frz). 24.20. Show that the tensor product of two unitary representations is unitary. 24.21. Switch the dummyindices of the double sum in (24.45), add (subtract)the two sums, and use (T ® T)ia.jb(g) = (T ® T)ai.bj(g) to show that the double sum can be written as a sum over the symmetric (antisymmetric) vectors alone. 24.22. Show that the characters XS(g) and XA(g) of the symmetrized and anti-

symmetrized productrepresentationsare giveu, respectively, by and 24.23. Supposethat

Show that

Aia ) transformsaccordingto T(a) , and A7) accordingto T(P) .

At)A7) transformsaccordingto T(axp).

24.24. Show that

Onecaninterpretthis as thestatementthatthe squareof thereducedmatrixelement is proportionalto theaverage(overi and j) of the squareofthefull matrixelements. 24.25. Constructthecharactertableof 84usingthe analyticalmethodandEquation

(24.57). 24.26. Find all the standardYoung tableauxfor 8s. Thus,determinethe dimension

of each irreducible represeutations of 8s. Check that the dimensions satisfy the second equationof (24.18). 24.27. Verify the remainingentries of Table 24.6. 24.28. Constructthe charactertable of 8s.

726

24. GROUP REPRESENTATION THEORY

24.29. Suppose that Q, an element of the group algebra of 5n , is given by n1

Q=

L

f

nj

1t i l

i=1

Show that and

_ 'Q . Q2 -no

24.30. Show that y(n)y(l") = O. Hint: There are as many even permutations in 5n as there are odd permutations. 24.31. Show that that the product of any two Young operators of 53 corresponding to different Young tableaux is zero and that (2, I ) y (2.1) _ 3y(2.1) Y1 1 1 '

(2,I)y (2,1) _ 3y(2,1) Y2 2 2 .

24.32. Construct the ideal .c~2,1) generated by dimensional.

and verify that it is two

y?,I)

24.33. Using the ideal.c\2,1) generated by y?,I), find the matrices of the irreducible representation T(2,1). From these matrices calculate the simple characters of 53 and compare your result with Table 24.4. Show that the ideal.c~2, I) generated by 2,1) gives the same set of characters.

yJ

24.34. Find all the Young operators for 54 corresponding to the first entry of 2

each row of Figure 24.1. Find the ideals .c\3,1) and .c\2 ) generated by the Young operators y?, I) and y?2) corresponding to the second and third rows of the table. 2

Show that .c\3,I) and .c\2 ) have 3 and 2 dimensions, respectively. 24.35. Verify the products of the Young frames of Figure 24.2.

Additional Reading 1, Barut, A. and Raczka, R. Theory ofGroup Representations andApplications, World Scientific, 1986. A comprehensive treatise on group theory and its application to physics written in a formal, but readable, language. 2. Boerner, H. Representation of Groups, North-Holland, 1963. A detailed introduction to group representation using group algebra techniques. 3. Elliott, J. and Dawber, P. Symmetry in Physics (2 volumes), Oxford University Press, 1979. Another informal book written for physicists.

24.9 PROBLEMS

727

4. Fulton, W. and Harris, J. Representation Theory, Springer-Verlag, 1991. A formal introduction to representation theory designed for graduate students in mathematics, but also useful for physicists with a solid math background. 5. Hamermesh, M. Group Theory and its Application to Physical Problems, Dover, 1989. A detailed and very readable account of the representation of syrmuetric groups written for physicists.

25

_

Algebra of Tensors

Einstein's summation convention

Up until a mere two decades ago, tensors were almost completely synonymous with (general) relativity--except for a minor use in hydrodynamics. Students of physics did not need to study tensors until they took a course in the general theory of relativity. Then they would read the introductory chapter on tensor algebra and analysis, solve a few problems to condition themselves for index "gymnastics," read through the book, learn some basic facts about relativity, and finally abandon it (unless they became relativists). Today, with the advent of gauge theories of fundamental particles, the realization that gauge fields are to be thought of as geometrical objects, and the widespread belief that all fundamental interactions (including gravity) are different manifestations of the same superforce, the picture has changed drastically. Two important developments have taken place as a consequence: Tensors have crept into other interactions besides gravity (such as the weak and strong nuclear interactions), and the geometrical (coordinate-independent) aspects of tensors have become more and more significant in the study of all interactions. The coordinateindependent study of tensors is the focus of the fascinating field of differential geometry and Lie groups, the subject of the remainder of the book. This chapter covers tensor algebra, while the next is devoted to tensor analysis. As is customary, we will consider only real vector spaces and abandon the Dirac bra and ket notation, whose implementation is most advantageous in unitary (complex) spaces. From here on, the basis vectors! of a vector space V will be distinguished by a subscript and those of its dual space by a superscript. If {ei }~1 is a basis in V, then {€j}f~! is a basis in V'. Also, Einstein's summation convention will be used:

1Wedenotevectors by roman boldface,andtensors of higherrank: by sans serif bold letters.

25.1 MULTILINEAR MAPPINGS

729

25.0.1. Box. Repeated indices, of which one is an upper and the other a k .

N

k'

lower index, are assumed to be summed over: a j bj means Li=1 a j hj.

As a result of this convention, it more natural to label the elements of a matrix representation of an operator A by af (rather than ai')' because then Ae, = afei'

25.1 Multilinear Mappings Since tensors arespecial kinds oflinear operators on vector spaces, let us reconsider t:..(V, W), the space of all linear mappings from the real vector space V to the real vector space W. We noted in Chapter 3 that t:.. (V, W) is isomorphic to a space with dimension dim V . dim W. The following proposition-whose proof we leave to the reader-shows this directly.

25.1.1. Proposition. Let{etl~l be a basis for V and (e~I:;"1 abasisforW. Then 1. the linear transformations T~ : V --> W in the vector space t:..(V, W), defined by (note the new way ofwriting the Kronecker delta)

j

=

1, ... , N1;

fJ

= 1, ... , N2,

form a basis in t:..(V, W). In particular, dim t:..(V, W) = NIN2. 2. If r

J are the matrix elements ofa linear transformation T E t:.. (V, W), then

T=

T:jT~.

The dual space V* is simply the space t:..(V, R), Proposition 25.1.1 (with N2 = I) then implies that dim V* = dim V, which was shown in Chapter 1. The dual

spaceis important inthediscussionof tensors, so we considersome of its properties below. The basis {T~I of Proposition 25.1.1 reduces to {T{I when W = lR and is denoted by {eiI7=1' with N = dimV* = dimV. The ei's have the property that . i e1 (e,) = 8, .

(25.1)

This relation was also established in Chapter 1. The basis B* = {eilf~l is simply the dual of the basis B = {e,1;;"1' Note the "natural" position of the indices for B and B*. Now suppose that {f,I;;"1 = B' is another basis of V and R is the (invertible) matrix carrying B onto B'. Let B '* = {'Pilf~l be the dual of B '. We want to find the matrix that carries B* onto B'*. If we denote this matrix by A and its elements we have by

af'

,k OJ

= epkfi

k i ,I k 1 ( )k = (k al e1)(r jie) j = al r j (J j = al r j = AR i '

730

25. ALGEBRA OF TENSORS

where the first equality follows from the duality of B' and B'*. In matrix form, this equation can be written as AR = 1, or A = R- I. Thus,

25.1.2. Box. The matrix that transforms bases of V* is the inverse of the

matrix that transforms the corresponding bases ofV.

In the above equations the upper index io matrix elements labels rows, and the lower iodex labels columns. This can be remembered by noting that the column vectors ei can be thought of as columns of a matrix, and the lower iodex i then labels those columns. Similarly, e J can be thought of as rows of a matrix. We now generalize the concept of linear functionals. multilinear mappings defined

25.1.3. Definition. A map T : VI x V2 X ... x V, ---> W is called r-linear if it is

linear in all its variables: ICVI,.", aVj +a'v~,

... , vr)

= aT(vlo.'" Vi, ...• Vr ) +a'T(Vl .... , v~, ... , v-)

for all i. We can easily construct a bilioear map. Let 71 E Vi and 72 E the map 71 ® 72 : VI x V2 ---> 1& by

Vi. We define (25.2)

The expression 71 ® 72 is called the tensor product of 71 and 72. Clearly, sioce 71 and 72 are separately lioear, so is 71 ® 72. An r-lioear map can be multiplied by a scalar, and two r-linear maps can be added; io each case the result is an r-lioear map. Thus, the set of r-lioear maps from VI x ... x V, ioto W forms a vector space that is denoted by .G(VI, ... , V,; W). We can also construct multilioear maps on the dual space. First, we note that we can define a naturallioear functional on V* as follows. We let 7 E V* and v E V; then 7(V) E 1&. Now we twist this around and define a mappiog v : V* ---> 1& given by V(7) sa 7(V). II is easily shown that this mappiog is lioear. Thus, we have naturally constructed a lioear functional on V* by identifyiog (V*)* with V. Construction of multilioear maps on V* is now trivial. For example, let VI E VI and V2 E V2 and define the tensor product VI ® V2 : Vi x Vi ---> 1& by (25.3) We can also construct mixed multilioear maps such as v ® 7 : V* x V ---> 1& given by v ® 7(0, u) = V(O)7(U) = O(V)7(U).

(25.4)

25.1 MULTILINEAR MAPPINGS

natural pairing of vectors and their duals

731

There is a bilinear map h : V* x V -+ lR that naturally pairs V and V*; it is givenby h(O, v) '" o(v). This mapping is called the natural pairing of V and V* into lR and is denotedby using angle brackets: h(O,v) '" (0, v) es O(v).

tensors; covariant and contravariant degrees

25.1.4. Definition. Let V be a vector space with dual space V*. Then a tensor of type (r, s) is a multilinear mapping 2 Tr

s

:

V* x V* x, ... X V* x V x V x ... x V -+ R I ,

,

r times

'

s times

The set of all such mappings for fixed rand s forms a vector space denoted by

Y,(V). The number r is called the contravariant degree of the tensor, and s is called the covariant degree ofthe tensor.

As an example,let VI,

... , Vr

E V and T I, ... , 1"8 E '\7*, and definethe tensor ® T' by

prodnct tj' es VI ® ... ® Vr ® T I ®

VI ® ... ® Vr ® T I ® ... ® T' (0 1,

, 0' , UI, ... , Us)

nn r

= VI(0 1) ... VT(Or)T I (UI)· .. T' (u,) =

s

0' (Vi)T j (uj),

i=l j=l

contravariant and covariant vectors and tensors tensors form an algebra multipiication of the algebra oftensors

Each v in the tensor product requires an element of V*; that is why the number of factors of V* in the Cartesian product equals the number of v's in the tensor product. As explainedin Chapter 0, the Cartesian product with s factors of V is denotedby V' (similarlyfor V*). A tensor of type (0, 0) is defined to be a scalar, so 'J1!(V) = R A tensorof type (1,0), an ordinary vector, is called a contravariant vector, and one of type (0, 1), a dual vector (or a linear functional),is called a covariant vector. A tensor of type (r, 0) is called a contravariant tensor of rank r, and one of type (0, s) is called a covariant tensor of rank s. The union of Y,(V) for all possible r and s can be madeinto an (infirtite-diroensional) algebra, called the algebra of tensors, by defining the following product on it: 25.1.5. Definition. The tensor product ofa tensorT oftype (r, s) and a tensor U of type (k, 1) is a tensor T ® U of type (r + k, s + 1), defined, as an operator on (V*yH x V,+I, by T ® U(OI ,

= T(OI,

, OTH, UI, , 0', UI,

, U,+I) , u,)U(OT+I, ... , r», U,+1, ... , U,+I).

This product turns the (infinite-dimensional) vector space of all tensors into an associative algebra called a tensor algebra. 2Just as the space '\1* of linear functionals of a vector space V is isomorphic to V, so is the space of tensors of type (T, s) to the tensor product space Vr,s (see Example 1.3.19).In fact, ~ (V) = Vi,s. as shownin [Warner, 83] on p. 58.

732

25. ALGEBRA OF TENSORS

This definition is a generalization of Equations (25.2), (25.3), and (25.4). It is easily verified that the tensor product is associative and distributive (over tensor addition), but not commutative. Making computations with tensors reqnires choosing a basis for V and one for V' and representing the tensors in terms of numbers (components). This process is not, of course, new. Linear operators arerepresented by arrays of nwnbers, i.e., matrices. The case of tensors is merely a generalization of that of linear operators and can be stated as follows:

25.1.6. Theorem. Let {eilf'::l be a basis in V, and (.ijJ=l a basis in V', usually

taken to be the dual of{eilf=l' Then the set ofall tensor products eil @ ... @ei, @ .h @ ... @.i, forms a basisfor T; (\1). Furthermore, the components ofany tensor A E T;(V) are i,· ) A ihl ......"is, = A( €h , ... ,€ ,eil"'" ei s .

Proof The proofis a simple exercise in employing the definition of tensor products and keeping track of the multilinear property of tensors. Details are left for the

D

reoo~

A useful result of the theorem is the relation (25.5) Note that for every factor in the basis vectors ofT; (V) there are N possibilities, each one giving a new linearly independent vector of the basis. Thus, the number of possible tensor products is N r +" and we have dim T; (V) = (dim VY+' .

25.1.7. Example. Let us consider the special case of '11 (V) as an illnstration. We can writeA E 'J'} (\7) as A = A~ei ® e j . Givenanyv E V, we obtain2 A(v) = (A~ei 0 ·.i)(v)

= A~ei r.i (v)]. '-v-'

ER

'11 (V),I:(V), and qv') are allthe same

This showsthat A(v) E V and Acan be interpretedas a linear operator on V, i.e., A E qV). Similarly, for T E \7* we get A(r)

= (A~ei 0ei)(r) = A~ [ei(r)]e i. '-v-'

ER Thus,A E L(V*). Wehaveshownthat givenA E 'JJ (V), there corresponds a linear operator belonging to I:(V) [or qv')1having a natnral retation to A.Similarty,given any A E I:(V) [or I:(V")] with a matrix representation in the basis {eilf'::l of V (or{ei lJ=l of V*) given by A~, thencorresponding to it in a natural way is a tensorin namely A~ei ® ei . Problem 25.5 shows that the tensor defined in this way is basis-independent. Therefore,

'JfcV),

25.1 MULTILINEAR MAPPINGS

natural isomorphism

733

TI

thereis a natural one-to-one correspondence among (V), C(V), and C(V*). Thisnatural correspondence is calleda naturalisomorphism. Whenever there is a natural isomorphism between two vector spaces, those vector spaces can be treatedas being the same.

III

We have defined tensors as multilinear machines that take in a vector from a specific Cartesian product space of V's and V*'s and manufacture a real number. Given the representation in Equation (25.5), however, we can generalize the interpretation of a tensor as a linear machine so that it takes a vector belonging to a Cartesian product space and manufactures a tensor. This corresponds to a situation in which not all factors of (25.5) find "partners." An illustration of this situation was presented in the preceding example. To clarify this let us consider A E (V), given by

Ti

A = A~kei <81 €i <81 €k.

ThismachineneedsaCartesian-productvectoroftheform(T, VI, V2), with-r E V* and VI, V2 E V, to give a real number. However, if it is not fed enough, it will not complete its job. For instance, if we feed it ouly a dual vector T E V*, it will give a tensor belonging to
If we feed it a double vector (VI, V2), it will manufacture a vector in V:

positioning 01 vectors and Iheir duals 10 match the tensor

What if we feed it a single vector v? It will blow its whistles and buzz its buzzers, because it does not know whether to give v to €i or €k (it is smart enough to know that it cannot give v to e.), That is why we have to inform the machine with which factor of e to match v. This is done by properly positioning v inside a pair of parentheses: If we write (, v, .), the machine will know that v belongs to €i, and (., "v) tells the machine to pair v with €k, If We write (v,., .), the machine will

give us an"error message" because it cannot pair v withej! The components of a tensor A, as given in Equation (25.5), depend on the basis in which they are described. If the basis is changed, the components change. The relation between components of a tensor in different bases is called the transformation law for that particular tensor, Let us investigate this concept. We use overbars to distiognish among various bases. For instance, B = Iei J~I'

B = leiJf~l' and B = lijklf=1 are three different bases of V. Similarly, B* = IdJ~l' B* = rEiJf=I' and B = l~kJf~1 are bases of V*. The components are also distioguished with overbars. Recall that if R is the matrix connectiog B and

2Here, we areassuming that Aactson anobject(suchasv) by ''pairing it up"with anappropriate factor of which Ais composed (such as ~.i).

734

25. ALGEBRA OF TENSORS

B, then

S = R- I connecls

B* and B*. For a tensor A of type (1,2), Theorem

25.1.6 gives

-Aijk=A\€ rxi - - ) ,ej,ek i

i m n p) = A(Sm€ ,Tjen,Tkep

n PA( em ,en,e ) =SmTjTk i n PAm np : p

(25.6)

=Smrjrk

This is the law that transforms the components of A from one basis to another. In the classical coordinate-dependent treatment of tensors, Eqnation (25.6) was the defining relation for a tensor of type (1,2). In other words, a tensor of type (1, 2) was defined to be a collection of nnmbers, A;:'p that transformed to another collection of nnmbers A~k according to the mIe in (25.6) when the basis was changed. In the modem treatment of tensors it is not necessary to introdnce any basis to define tensors. Only when the components of a tensor are needed mnst bases be introduced. The advantage of the modem treatment is obvious, since a (1, 2)-type tensor has 27 components in three dimensions and 64 components in four dimensions, and all of these are represented by the single symbol A. However, the role of components should not be downplayed. After all, when it comes to actual calcnlations, we are forced to choose a basis and manipulate components. Since Y,(V) are vector spaces, it is desirable to investigate mappings from one snch space 10 another. We will be particularly interested in linear mappings. For example, f : ~J(V) --> 'Ji!(V) = JR is what was called a linear functional before. Similarly, t : ~J(V) --> ~J(V) is a linear transformation on V. A special linear transformation is tr : ~l (V) --> 'Ji!(V) = JR, given by tr A = tr(A~ei ®

N

fj)

sa A: sa ~::.>:. ;=1

This is the same trace encountered in the study oflinear transformations in Chapter

3. Although the above definition makes explicit use ofcomponents with respect to a basis, it can be easily shown (see Problem 25.7) that it is in fact basis-independent. Functions of tensors that do not depend on bases are called invariants. Another example of an invatiant is a linear functional (see Problem 25.6). 25.1.8. Example. Consider the tensorA E T5(V) givenby A = "I 0 "I

+ "2 0

"I.

We

calculate the analogue of the trace for A: LT=l Aii = 1 + 0 = 1. Now we change to a new basis, {Cl, ez}. given by ei = Cl + 2cz ande2 = -cl + f2. In terms of the new basis vectors, A is given by

A = (01 + 2ez) 0 (01 + 2ez) =3°2001

with L:T=l

+ 6°2 0

Aii = 0 + 6 =

+ (-01 + 02) 0

(01 + 2°2)

°2

6 #:- L:f=l Aij. This kind of "trace" is not invariant.

II

25.1 MULTILINEAR MAPPINGS

735

Whenever a quantity is defined without reference to a basis, it is clearly invariant. The trace is an example of an invariant quantity whose definition does require a basis. Another example is the determinant of a linear operator [or a tensor of type (I, 1)], in terms of a matrix representation of that operator, which is clearly basis-dependent. A basis-independent definition for the determinant of a tensor will be given later in this chapter. Besides mappings of the form h : 'Y, (V) ->(V) that depend on a single variable, we can define mappings that depend on several variables, in other words, that take several elements of 'Y, (V) and give an element of 'Jj' (V). We then write h : ('Y,(v))m ->- 'Jj'(V). It is understood that h(tl, ... , tm ) , in which all ti are in 'Y,(V), is a tensor of type (k,I). If h is linear in all of its variables, it is called a tensor-valned multilioear map. Furthermore, if h(tl, ... , tm ) does not depend on the choice of a basis of 'Y, (V), it is called a multilinear invariant. In most cases k = 0 = I, and we speak of scalar-valued invariants, or siroply invariaots. An example of a multilinear invariant is the determinant considered as a function of the rows of a matrix. The following defines an important class of multilinear invariants:

'If

tensor-valued multilinear map

contraction of a tensor

'Y, (V) with respect to a contravariant index p and a covariant index q is a linear mapping C : 'Y, (V) ->'J;=!(V) given in components by 25.1.9. Definition. A contraction of a tensor A E

It can be readily shown that contractions are invariants. The proof is exactly the same as that for the invariance of the trace. In fact, the trace is a special case

of a contraction, in which r = s = 1. The linearity inherent in the construction of tensor algebras carries along some of the properties and structures of the underlying vector spaces. One such property is isomorphism. Suppose that F : V ->- U is a vector space isomorphism. Then F* : U* ->- V*, the pullback of F, is also an isomorphism (Proposition 1.3.18). Associated to F is a linear map-which we denote by the same symbol-from 'Y, (V) to 'Y, (U) defined by

where T E 'Y, (V), (Ji E U*, and U j E U. The reader may check that tltis map is an algebra isomorphism (see Section 1.4). We shall use tltis isomorhism to define

derivatives for tensors in thenext chapter.

736

25. ALGEBRA OF TENSORS

25.2 Symmetries of Tensors Many applications demand tensors that have some kind of symmetry property. We have already encountered a symmetric tensor-the metric "tensor" of an inner product: If V is a vector space and VI, V2 E V, then g(VI, V2) = g(V2, VI). The following generalizes this property. symmetric tensor defined

25.2.1. Definition. A tensor A is symmetric in the i th and j th variables if its value as a multilinear function is unchanged when these variables are interchanged. Clearly, the two variables must be ofthe same kind. An immediate consequence of this definition is that in any basis, the components of a symmetric tensor do not change when the ith and jth indices are interchanged.

contravariantsymmetric; covarlant-svrnmetrlc; symmetric

25.2.2. Definition. A tensor is contravariant-symmetric if it is symmetric in every

pair ofits contravariantindicesand covariant-symmetric if it is symmetricin every pair of its covariant indices. A tensor is symmetric symmetric and covariant-symmetric.

if it

is both contravariant-

The set of all symmetric tensors S' (V) of type (r, 0) forms a subspace of the vector space'' TQ. Similarly, the set of symmetric tensors of type (0, s) forms a subspace S, of 'J;>. The (independent) components of a symmetric tensor A E S' are Ailiz...ir. where it ::: iz ~ ... ::: i-; the other components are given by symmetry. Although a set of symmetric tensors forms a vector space, it does not form an algebra under the usual multiplication of tensors. In fact, even if A = Aije, @ej and B = Rk1ek @ el are symmetric tensors of type (2,0), the tensor product A@ B = Aij Rkl ei @ej @ek @el need not be a type (4, 0) symmetric tensor. For instance, A i k s!' may not equal Aij R kl. However, we can modify the definition of the tensor product (for symmetric tensors) to give a symmetric product out of symmetric factors. symmetrizer

25.2.3. Definition. A symmetrizer is an operator § : TQ

""* S' given by (25.8)

where the sum is taken over the r! permutations a/the integers 1,2, ... , r, and r l , ... , r' are all in V'. §(A) is often denoted by A,. Clearly, A, is a symmetric tensor.

A similar definition gives the symmetrizer § : 'J;> in (25.8), we would have VI, ... , V,. 3When there is no risk of confusion, we shall delete V from

given underlyingvector space.

""* S,. Instead of r l , ... , r'

'.J; (V), it being understood that all tensors

are defined on some

25.2 SYMMETRIES OF TENSORS

737

25.2.4. Example. For r = 2, we have only two permutations,and

Forr = 3, we havesix permutations of 1, 2, 3, and(25.8) gives As(r I, r 2 , r 3 )

A(r 2, r I, r 3 ) + A(r I, r 3 , r 2 ) + A(r 3, r I, r 2 ) + A(r3, r 2 , r I) + A(r2, r 3 , rI)]

= !lA(rI, r 2 , r 3 ) +

It is clearthatinterchanging any pairof r's on the RHS of the above two equations does

not change the sum. Thus, As is indeed a symmetric tensor.

II

It can be shown that

dimSr(V) =

(N +rr -

1) sa

(Nrl(N+ r- -1)1. 1)1

The proof involves counting the number of different integers iI, ... , i; for which 1:::; im ~ im+ l ~ N for each m. We are now ready to define a product on the collection of synnnetric tensors and make it an algebra, called the symmetric algebra. symmetric tensors form an algebra under thisproduct

25.2.5. Definition. The symmetric product of symmetric tensors A E sr (V) and B E S' (V) is denoted by AB and defined as

r+, _ (r

1

AB(T , ... , T ) =

+ s)!",,(A<8IB)(T1 , ... ,Tr+, ) rJs1

= _1_ TIs!

LA (T"(I), ... , T,,(r») B (T"(r+I), ... , T,,(r+,») , 1r

where the sum is over all permutations of 1, 2, ... , r of A E s, (V) and B E S, (V) is defined similarly.

+ s. The symmetric product

25.2.6. Example. Let us construct the symmetric tensor products of vectors. First we find the symmetricproduct of vj and v2 both belonging to V = (V):

'J6

(VIV2)(rI,~)

== VI (rI)V2(~) +

VI (~)V2(rI)

= VI(r I)V2(T2) +V2(r I)VI(T2) = (VI 0 v2 + v2 0 vI)(r I, r ) . 2

Sincethisis true for any pair T 1 andT 2, we have

In general, VI vz

... v r =

En VJr(l) ® Vn(2) ® .'. . ® V1r (r ) '

iii

738

25. ALGEBRA OF TENSORS

It is clear from the definition that symmetric multiplication is commutative, associative, and distributive. If we choose a basis {eilr:,l for V and express all symmetric tensors in terms of symmetric products ofej using the above properties, then any symmetric tensor can be expressed as a linear combination of terms of the form (ell n l ••• (eNinN. Skew-symmetry is the same as symmetry except that in the interchange of variables the tensor changes sign. The following theorem, whose simple proof we leave as a problem, further characterizes skew-symmetric tensors. skew-symmetric

tensors

25.2.7. Theorem. A tensor A is skew-symmetric in contravariant indices i and j if and only if for all T E V*, substitution of T E V* for both the i th and j th variables of A yields zero, regardless of the values of the remaining variables.

covariant and contravariant skew-symmetric

tensors

25.2.8. Definition. A covariant (contravariant) skew-symmetric tensor is one that is skew-symmetric in all pairs of covariant (contravariant) variables. A tensor is skew-symmetric if it is both covariant and contravariant skew-symmetric.

Leopold Kronecker (1823-1891) was the son of Isidor Kronecker, a businessman, and Johanna Prausnitzer;: They were wealthy andprovided private tutoring athomefor their son until he entered the Liegnitz Gymnasium. At the gymnasium, Kro-

necker's mathematics teacher was E. E. Kummer, who earlyrecognizedtheboy's ability andencouraged himto do independent research. He also received Evangelical religiousinstruction, although he was Jewish; he formally converted to Christianity in the lastyearof his life. Kronecker matriculated attheUniversity of Berlinin 1841, where he attended lectures in mathematics givenby Dirichlet. LikeGauss andJacobi, he was interested in classical philology. He also attended Schelling's philosophy lectures; he was later to makea thorough studyof theworks of Descartes, Spinoza, Leibniz, Kant, andHegel, as well as those of Schopenhauer, whoseideas he rejected. Kronecker spentthe summer semester of 1843 at the University of Bonn, andthe fall semester atBreslau (nowWroclaw, Poland) becauseKummer hadbeenappointed professor there. He remained there for two semesters, returning to Berlinin thewinter of 1844--1845 to take thedoctorate. Kronecker tookhis oralexamination consistingof questions not only in mathematics, butalsoin Greek historyof legal philosophy. He wasawarded thedoctorate on IOSeptember 1845. Dirichlet, hisprofessor andexaminer, wasto remain one of Kronecker's closestfriends, as was Kummer, his first mathematics teacher. In the meantime, in Berlin, Kronecker was also becomingbetter acquainted with Eisenstein and with Jacobi. During the sameperiod Dirichlet introduced himto Alexander von Humboldt andto thecomposer Felix Mendelssohn, who was bothDirichlet's brother-in-law andthe cousinof Kummer's wife. Familybusiness thencalled Kronecker from Berlin. In its interest he was required to spend a few yearsmanaging anestatenearLiegnitz, as well as to dissolve thebanking business of anuncle. In 1848he married thelatter's daughter, his cousinFanny Prausnitzer; they

25.3 EXTERIOR ALGEBRA

739

had six children. Having temporarilyrenounced an academic career, Kroneckercontinued to do mathematics as arecreation. He both carried on independent research and engaged in a lively mathematical correspondence with Kummer; he was not ambitious for fame, and was able to enjoy mathematics as a true amateur. By 1855,however, Kronecker's circwnstances had changed enough to allow him to return to the academic life in Berlin as a financially

independentprivate scholar. In 1860 Kummer, seconded by Borchardt and Weierstrass, nominated Kronecker to the Berlin Academy, of which he became full member on 23 January 1861. Kronecker was increasingly active and influential in the affairs of the Academy, particularly in recruiting the most important German and foreign mathematicians for it. His influence outside Germany also increased. He was a member of many learned societies, among them the Paris Academy and the Royal Society of London. He established other contacts with foreign scientists in his numerous travels abroad and in extending to them the hospitality of his Berlin home. For this reason his advice was often solicited in regard to filling mathematical professorships both in Germany and elsewhere; his recommendations were probably as significant as those of his erstwhile' friend Weierstrass. The cause of the growing estrangement between Kronecker and Weierstrass was partly due to the very different temperaments of the two, and their professional and scientific differences. Since they had long maintained the same circle of friends, their friends, too, became involved on both levels. A characteristic incident occurred at the new year of 18841885, when H. A. SChwarz, who was both Weierstrass's student and Kummer's son-in-law, sent Kronecker a greeting that included the phrase: "He who does not honor the Smaller [Kronecker], is not worthy of the Greater [Weierstrass]." Kronecker read this allusion to physical size-he was a small man, and increasingly self-conscious with age-as a slur on his intellectual powers and broke with Schwarz completely. Kronecker's mathematics lacked a systematic theoretical basis. Nevertheless, he was preeminent in uniting the separate mathematical disciplines. Moreover, in certain ways-his refusal to recognize an actual infinity, his insistence that a mathematical concept must be defined in a finite number of steps, and his opposition to the work of Cantor and Dedekindhis approach may be compared to that of intuitionists in the twentieth century. Kronecker's mathematics thus remains influential.

25.3 Exterior Algebra The following discussion will concentrate on tensors of type (r, 0). However, interchanging the roles of V and V* makes all definitions, theorems, propositions, and conclusions valid for tensors of type (0, s) as well. The set ofall skew-symmetric tensors oftype (p, 0) forms asubspace of Tr'(V). This subspace is denoted by AP(V). His not, however, an algebra unless we define a skew-symmetric product analogous 10 that for the symmetric case. First, we need the following: anlisymmelrizer

25.3.1. Definition. Anantisymmetrizerisalinearoperatoril.: ~C(V) -> AP(V),

740

25. ALGEBRA OF TENSORS

given by (25.9)

The sum is over all permutations oj (12 ... pi, and E" = E"CI)... ,,(p) is + I if the permutation is even and -I if the permutation is odd. A(T) is sometimes denoted byTa. 25.3.2. Example. Let us write out Equation (25.9) for p = 3. The procedure is entirely analogousto Example 25.2.4: Ta(7"I, 7"2, 7"3) = ~[€123A(7"I, 7"2, 7"3)

+ €213A(7"2, 7"1,7"3) + E132A(7"I, 7"3,7"2)

+ €312 A(7"3, 7"1,7"2) + €32I A(7"3, 7"2, 7"1) + E23I A(7"2, 7"3, 7"1)] = ~[A(7"I, 7"2, 7"3) _ A(7"2, 7"1,7"3) _ A(7"I, 7"3,7"2)

+ A(7"3, 7"1,7"2) _

A(7"3, 7"2, 7"1) + A(7"2, 7"3, 7"1)].

Thereader may easilyverifythatallterms withaplussignareobtained from (123) byaneven number of interchanges of symbols, andthose with a minus signby an oddnumber. II

We can now define the important product that turns the collection of all antisymmetric tensors into an algebra. . exterior product defined

25.3.3. Definition. The exteriorproduct (also called the wedge, Grassmann, alternating, or veckproduct) oJtwoskew-symmetric tensors A E AP(V)andB E Aq(V) is a skew-symmetric tensor belonging to Ap+q (V) and given bl A /\ B es (r

+ s)! A(A ® B) =

(r

r!s!

+ s)! (A ® B)a.

r!s!

More explicitly,

25.3.4. Example. Let us find the exterior product of vectors. Firstwe findthe exterior product of VI aod vz both belonging to V = TJ(V): VI!\ V2(7"I, 7"2)

=

L E,,(I),,(2)(VI ® V2) (7""(1), 7""(2)) x

= E12(VI ® V2)(7"I, 7"2) + E21(VI ® V2)(7"2, 7"1)

= VI (7"I)V2(7"2) -

VI(7"2)V2(7"I)

= (VI ® V2 - V2 ® VI)(7"I,

,-2).

4Thereader shouldbe warned that different authors mayuse different numerical factors in thedefinition of theexterior product.

25.3 EXTERIOR ALGEBRA

741

Since this is true for arbitrary T 1 and T 2, we have

In general, vI A VZ A ... A Vr

= I:',,(l)...,,(r) V,,(l) ® V,,(Z) ® ... ® V,,(r)'

"

In particular, this shows that the exterior product (of vectors) is associative.

III

If {ejI7=1 is a basis with dual {fil~l' then the lasiresult of Exarnple 25.3.4 yields (25.10) The following theorem contains the properties of the exterior product (for a proof, see [Abra 88, p. 394]): 25.3.5. Theorem. The exteriorproduct is associative and distributive with respect to the addition oftensors. Furthermore, it satisfies thefollowing anticommutativity property:

whenever A E AP (V) and B E A q (V). In particular, VI A Vz = -Vz A VI for VI,VZ E

V.

The wedge product of linear functiouals of a vector space is particularly important in the analysis of tensors, as we shall see in the next chapter. p-forms defined

pull-back of p-forms bylinear transformations

25.3.6. Definition. The elements of AP(V*) are called p-forms. A linear transformation T ; V ~ W induces a transformation'' T* AP(W*) ~ AP(V*) defined by

P

E

AP(W*), Vi E V.

(25.11) T*P is called the pullback of p by T. The most important properties of pullback maps are given in the following:

25.3.7. Proposition. Let T : V ~ U and S : U I. T* : AP(U*) 2. (S

0

~

~

W. Then

AP(V*) is linear. (So is S*, ofcourse.)

T)* = T* 0 So.

6Note that T* is the extension of the pullback operator introduced at the end of Chapter 1.

742

25. ALGEBRA OF TENSORS

3. 1fT is the identity map, so is T*. 4. 1fT is an isomorphism, so is T* and (T*)-l = (T- 1)*.

5. If P E AP(W*) and 0-

E

Aq(W*), then T*(p

A

0") = T*p

A

T*o-.

Proof The proof follows directly from definitions and is left as an exercise for the reader. D The components of A E AP(V) are given by Ai, ...i p , where i1 < i2 < ... < i p • All other components are related to these by skew-symmetry. The number of independent components, which is the dimension of A P (V), is equal to the number of ways p numbers can be chosen from among N distinct numbers in such a way that no two of them are equal. This is simply the combination of N objects taken p at a time. Thus, we have dim AP(V)

= (N) = P

N! . p!(N - p)!

(25.12)

In particular, dim AN (V) = 1. This should come as no surprise, because from a basis {eil!:l of V, we can form a basis for AP(V) by constructing all products of the Iorm e., Aei2 A·· ·Aeip [of which there are (~)]. However, whenp = N, thereis, within a multiplicative constant, only one such product and that is ei Ae2A· .. AeN. An elegant way of determining the linear independence of vectors using the formalism developed so far is given in the following proposition. 25.3.8. Proposition. A set of vectors, VI, ... , v p, is linearly independent only if VI A .. · A v p 01 o.

if and

Proof If {Vi J;=l are independent, then they span a p-dimensional subspace M of V. Considering M as a vector space in its own right, we have dim A P (M) = 1. A basis for AP(M) is simply VI A ... A v p , which cannot be zero. Conversely, suppose that "'1VI + ... + "'p v P = O. Then taking the exterior prodnct of the LHS with V2 A V3 A ... A v P makes all terms vanish (because each will have two factors of a vector in the wedge product) except the first one. Thns, we have "'1VI A ... A v P = O. The fact that the wedge product is not zero forces "'1 to be zero. Similarly, mnltiplying by VI A V3 A ... A v P shows that "'2 = 0, and

D

00=

25.3.9. Example. Let{eil!:l be a basis for V. Let vI 2e3, v3 = -e1 - 3e2 + 2e3. Take the wedge product of the firsttwo v's: VI A V2

= (e1 + 2e2 -

= ei +202 -

e3. V2

= 3e1+e2 +

e3) A (3e1 + ez + 2e3) = -5e1 Ae2 + 5e1 A e3 + 5e2 A e3.

All the wedge products that haverepeated factors vanish. Now we multiply by v3: VI AV2 AV3

= -5e1 Ae2 A (-e1 -

3e2 +2e3) +5e1 Ae3 A (-e1 - 3e2 +2e3)

+ 5e2 A e3 A (-e1 - 3e2 + 2e3) = -WeI A eZ A e3 - 15el A e3 1\ eZ - 5eZ A e3 A et

= O.

25.3 EXTERIOR ALGEBRA

Weconcludethat the three vectorsare linearlydependent.

743

III

As an application of Proposition 25.3.8, let us prove the following. Cartan's lemma

25.3.10. Theorem. (Cartan's lemma) Suppose that lei };~1' P ::: dim V, form a linearly independent set ofvectors in V and that IVilf~1 are also vectors in V such that L;~1 ei /\ Vi = O. Then all Vi are linear combinations of only the set lei };~1' Furthermore,

if Vi = L}=1 Aijej, then Aij = Aji.

Proof Multiplying both sides of L;=1 ei /\ Vi = 0 by ez /\ ... /\ e p gives -VI /\ er /\ e2 /\ ... /\

ep = O.

By Proposition 25.3.8, VI and the ei are linearly dependent. Similarly, by multiplying L;~1 ei /\ Vi = 0 by the wedge product with ek missing, we show that Vk and the ei are linearly dependent. Thus, Vk = L;=1 Akiei, for all k. Furthermore,

we have p

o=

p

I:ek /\ Vk = L k=l

p

I:ek /\ (Akiei) = L(Aki - Aik)ek /\ er,

k=l i=l

k-ci

where the last sum is over both k and i with k < i, Clearly, Iek /\ ei } with k < i are linearly independent (show this!). Therefore, their coefficients must vanish. D

Elie Joseph Carlan (1869-1951), born in Dolomieu (near Chambery), Savoie, Rhone-Alpes, France, became a student at the Ecole Normale in 1888 and obtained his doctorate in 1894. He lectured at Montpellier (1894-1896), Lyon (18961903), Nancy (1903-1909), and Paris (1909-1940). He had fourchildren, one of whom,Henri Cartan, was to producebrilliantwork in mathematics. Two othersdied tragically. Jean,a composer, died atthe age of25, while Louis,a physicist, was arrested by theGermans in 1942andexecuted after 15months in captivity. Cartan added greatly to the theory of continuous groups, which had been initiated by Lie. His thesis (1894) contains a major contribution to Lie algebras wherein he completed theclassification of thesemisimple algebras that Killing had essentially found. He then turned to thetheory of associative algebras andinvestigated the structure for these algebras overthereal and complex fields. Wedderburn wouldcomplete Cartan's work in thisarea. He then turned to representations of semisimple Lie groups. His work is a striking synthesis of Lie theory, classical geometry, differential geometry, and topology, whichwas tobe found in all Cartan's work. Healsoapplied Grassmann algebra tothetheory of exterior

differential forms, By 1904 Cartan was turning to papers on differential equations, andfrom 1916 on he published mainly on differential geometry. Klein's Erlanger Program was seen to be

744

25. ALGEBRA OF TENSORS

inadequateas a geueraldescriptionof geometryby Weyl and Veblen, and Carlanwasto play a major role. He examineda space acted on by an arbitrary Lie group of transformations, developing a theory of movingframes that generalizes the kinematical theory of Darboux. Cartan further contributed to geometry withhis theory of symmetric spaces, whichhave their origins in papers he wrote in 1926.Itdevelops ideasfirst studied by Clifford and Cayley

and used topologicalmethods developedby Weyl in 1925. This work was completedby 1932. Cartan then went on to examineproblems on a topic first studied by Poincare. By this stage his son, Henri Cartan,wasmaking major contributions tomathematics, and ElieCartan was ableto buildon theorems proved by his son. Cartan also published work on relativity andthe theory of spinors. He is certainly one of themost important mathematicians of the first half of thetwentieth century.

25.3.1 The Determinant One of the most beautiful applications of exterior algebra is in the theory of determinants. We have already considered determinants in detail in Chapter 3, where we noted how messy it was to prove some of the theorems. With the machinery of exterior algebra at our disposal, we will see how elegant this theory becomes and how trivial some of the proofs will tum out to be. First, let us recall that dim AN (V) = I when N = dim V. This means that if {ed!:! is a basis of V, then ej /\ez/\' . ·/\eN is the only vector in the corresponding basis of A N (V). On the other hand, if {Vi I!:! is any set of N vectors, the product vi /\ Vz /\ ... /\ VN is either zero (if the Vi'S are linearly dependent) or a nonzero /\ eN is a basis of fIN (V), we product belonging to AN (V). Since e! /\ ez /\ conclude that for any set of N vectors, vi. vz. , VN E V, the product vi /\ Vz /\ ... /\ VN is a multiple (possibly zero) of e! /\ ez /\ ... /\ eN. Now let A be a linear operator on V. The set of vectors Ae!, Aez,... , AeN all belong to V. By the above remarks (Ae!) /\ ... /\ (AeN) is proportional to e! /\ ez /\ ... /\ eN. We now show that the proportionality constant is simply det A. determinant in terms ofwedge products

25.3.11. Theorem. Let A: V --> V be linear. Let {ed!:! be a basisfor V. Then (Ae!) /\ ... /\ (AeN) = (detA)e! /\ ... /\ eN.

Proof Let Aer

= A~eir' for r

=

(25.13)

1,2, ... , N. Then

It is straightforward to show that (25.14)

25.3 EXTERIOR ALGEBRA

745

so that (Ael) /\ ... /\ (AeN) = (A\l A~ ... A~ Eili2...iN)ej /\ ez /\ ... /\ eN. The expression in parentheses on the RHS is simply the determinant as defined in Chapter 3 (recall the summation convention). D Levi-Civita tensor

25.3.12. Example. The symbol Eili2...iN'calledthe Levi-Civita tensor, can be defined

and determinants

by Equation (25.14). Consider the linear operator E whose actionon a basis {ed~l is to permute the vectors so that Eek = eik' Then on theone hand,

andon theother hand, by thedefinition of theaction of E, (Eel) A ... A (EeN)

= eit

1\ ei2 A ... 1\

eiN'

Comparison of these two equations with (25.14) gives detE = filiz ...iN"

III

Since the determinant is hasis-independent, the result of the previous example can be summarized as follows: 25.3.13. Box. The Levi-Civita tensor Eili2 ...iN takes the same value in all coordinate systems.

Let us now look at the determinant of the product of two operators A and B. By definition, (ABel) /\ ... /\ (ABeN) = [det(AB)]ej /\ ... /\ eN. However, we also have (ABel) /\ ... /\ (ABeN) = [A(Bel)]/\'" /\ [A(BeN)] = (det A)(Bej) /\ ... /\ (BeN) = (detA)(det B)el /\ ... /\ eN. It follows that detAB = (det A)(det B). Here the power and elegance of exterior algehra can be truly appreciated-to prove the same result in Chapter 3, we had to go through a maze of index shuffling and reshuffling.

25.3.2 Orientation The reader is no doubt familiar with the right-handed and left-handed coordinate systems in ~3. In this section, we generalize the idea to arbitrary vector spaces. oriented basis defined

25.3.14. Definition. An oriented basis of an Nsdimensional vector space is an ordered collection of N linearly independent vectors. If {Vi l;;"j is one oriented basis and {uil;;"1 is a second one, then UI/\ U2/\'"

/\ UN

= (det Rjvj

A V2/\"'/\ VN,

where Ris the transformation matrix and det Ris anonzero number (R is invertible), which can be positive or negative. Accordingly, we have the following definition.

746

25. ALGEBRA OF TENSORS

oriented vector spacesdefined

25.3.15. Definition. An orientation is the collection ofall oriented bases related

by a transformation matrixhavinga positive determinant.A vector spacefor which an orientation is specified is called an oriented vector space. Clearly, there are only two orientations in any vector space. Each oriented basis is positively related to any oriented basis belonging to the same orientation and negatively related to any oriented basis belonging to the other orientation. For example, inJR3 , the bases {ex, ey, e,} and {ey, ex, ez} belong to different orientations

because

The first basis is (by convention) called a right-handed coordinate system, and the second is called a left-handed coordinate system. Any otherbasis is either right-handed or left-handed. There is no third altemative!

volume element ofa vector space

positive orientation

25.3.16. Definition. Let V be a vector space. Let V* have the oriented basis (e i }~I' The oriented volume element I-' E AN (V*) of V is defined as

Note that if [e.] is ordered as {ej } , then I-'(e" e2, ... , ev) = +liN!, and we say that {e.} is positively oriented with respect to 1-'. In general, {v;} is positively oriented with respect to I-' if I-'(VI, V2, ... , VN» O. The volume element of V is defined in terms of a basis for V*. The reason for this will become apparent later, when we shall see that dx, dy, and dz form a basis for (ffi.3)*, and dx dy dz sa dx A dy Adz.

25.3.3 Symplectic Vector Spaces Mechanics was a great contributor to the development of tensor analysis. It provided examples of manifolds that went beyond mere subspaces of JRn. The phase space of Hamiltonian mechanics is a paradigm of manifolds that are not "hypersurfaces" of some Euclidean space. We shall have more to say about such manifolds in Chapter 26. Here, we shall be content with the algebraic strocture underlying classical mechanics. symplectic form, symplectic vector space,and sympiectic transformation

25.3.17. Definition. A 2-form w E A 2 (V*) is called nondegenerate if w(v], V2) = 0 for all VI E V implies V2 = O. A symplectic form on V is a nondegenerate 2-form WE A 2(V*). The pair (V, w) is called a symplectk vector space. If (V, w) and (W, p) are symplectic vector spaces, a linear transformation T : V --+ W is called a symplectic transformation or a symplectic map if

T*p=w. Any 2-form (degenerate or nondegenerate) leads to other quantities that are also of interest. For instance, given any basis {Vi} in V, one defines the matrix of

25.3 EXTERIOR ALGEBRA

The flat, andsharp d maps defined

747

the 2-form WE A 2(V*) by wii sa W(Vi, vi)' Similarly, one can define the nseful V ---> V* by linear map

w' :

eR

'[W'(v)]Vi es w(v, Vi).

(25.15)

~

e11*

rank ofa symplectic form

The rank of w' is called the rank of w. The reader may check that

25.3.18. Box. A 2-form W is nondegenerate if and only if the determinant of (wi}) is nonzero, if and only if w' is an isomorphism, in which case the inverse ofw' is denoted by w~.

canonical basis ofa symplectic vector space

25.3.19. Proposition. Let V be an N-dimensional vector space and W E A 2(V*). Ifthe rankofw is r, thenr = 2nfor some integer n and there exists a basis [e.} ofV, calleda canonicalbasis of anda dual basis {Ei},suchthatw = L'i~1 Ei/\EJ+n, or, equivalently, the N x N matrix ofw is given by

V,

10) o 0 o

0

where 1 is the n x n identity matrix and 0 is the (N - 2n) x (N - 2n) zero matrix.

e;)

Proof Since w 01 0, there exist a pair of vectors ei, e; E V snch that w(el, 01 O. Dividing el by a constant, we can assume w(el, e;) = 1. Because of its antiLet VI symmetry, the matrix of w in the plane Pj spanned by er and e; is C~\ be the w-orthogonal complement of :PI, i.e.,

A).

VI

= (v E V I w(v, VI) = 0

V VI E :PJl.

Then the reader may check that P,

n VI = O. Moreover, V = :PI + VI because

v = ,w(v, e;)el - w(v, el)e;, +":' - w(v, e;)el

+ w(v, el)e;,

eVr (Reader, verify!)

for any v E V. Thus, V = :PI E9 VI. If w is zero on all pairs of vectors in VI, then we are done, and the rank of w is 2; otherwise, let e2, e; E VI be such that w(e2, e;) 01 O. Proceeding as above, we obtain VI

= :P2 E9 V2

=}

V

= :PI E9:P2 E9 V2,

where :P2is the plane spanned by 02, and e; and V2 its w-orthogonal complement in VI. Continuing this process yields V = :PI E9:P2 E9 ... E9:Pn E9 Vn,

748

25. ALGEBRA OF TENSORS

where V.. is the subspace of V ou which w is zero. This shows that the rank of w is 2n. By reordering thebasis vectors such that e~ == en+k, we construct a new basis lei 1f:,1 in which w has the desired matrix. To conclude the proposition, it is sufficient to show that L:J= l ei A e i +n, in which {eilf=, is dual 10 lei 1f:,1 ,has the same matrix asw. This is left as an exercise

D

furthereoo~

We note that in the canonical basis,

if i, j :::: n, if j = n + k, k ::: n,

ifi ::0: 2n or j ::0: 2n. If we write v E V as v = L7=, (Xiei + Yie..+i) + Lf:,j 2 .. ziC2..+i in the canonical basis of V, with a corresponding expansion for v', then the reader may verify that

.

w(v, v') = ~)xiyi - XiYi). i=l

The following proposition gives a useful criterion for nondegeneracy of w: 25.3.20. Proposition. Let w be a 2-form in the finite-dimensional vector space V. Then w is nondegenerate iff V has even dimension, say 2n, and w" es w A·· . A w is a volume element ofV.

Proof Suppose w is nondegenerate. Then, w' is an isomorphism. Therefore, the rank of w, an eveu number by Proposition 25.3.19, must equal dim 17* = dim V. Moreover, by taking successive powers of w and using mathematical induction, one can show that w" is proportional to e l A ... A e2", Conversely, if w n ex: e 1 1\ ... /\ e2n .is a volume element, then by Proposition 25.3.8, the {eil are linearly independent. Furthermore, dim 17* must equal the number of linearly independent factors in the wedge product of a volume element. Thus, dim 17* = 2n.But2n is also the rank ofw. Itfollows thatw' is onto. Since V is finite-dimeusioual, the dimension theorem implies that is an isomorphism. D

w'

25.3.21. Example. Let V be a vector space and 17* its dual. The direct sum V Ell '.1* can be turned into a symplecticvector space if we define W E A2(17 Ell '.1*) by W(v + 1", v'

+ 1"') == 1'" (v) -

I"(v),

where v, v' E V and 1",1'" E '.1*. The reader may verify that (V Ell '.1*, w) is a symplectic vectorspace. Thisconstruction of symplectic vectorspacesis closely related to Hamiltonian dynamics,to which we shall return in Chapter26. III

25.4 INNER PROOUCT REVISITEO

749

Suppose (V, w) and (W, p) are 2n-dimensional symplectic vector spaces. Then, by Proposition 25.3.7, any symplectic mapT : (V, w) --+ (W, p) is volumepreserving, i.e., (T*p)n is a volume element of W. It follows lbat lbe rank of T* is 2n, and by Proposition 1.3.18, so is lbe rank ofT. The dimension lbeorem now implies lbat T is an isomorphism. Symplectic transformations on a single vector space have an interesting property:

symplectic group

25.3.22. Proposition. Let (V, w) be a symplectic vector space. Then the set of symplectic mappings T : (V, w) --+ (V, w) forms a group under composition, called the symplectic group and denoted by Sp(V, w).

Proof Clearly, Sp(V, w) is asubsetofGL(V). One need only show lbatlbeinverse of a symplectic transformation is also such a transformation and lbat lbe product of two symplectic transformations is a symplectic transformation. The details are left for lbe reader. 0 symplectic matrices

A matrix is called symplectic ifit is lbe representation of a symplectic transformation in a canonical basis of lbe underlying symplectic vector space. The reader may check lbat lbe condition for a matrix A to be symplectic is At JA = J, where J is the representation of w in the canonical basis: J=

(0-1 01) '

where 1 and 0 are lbe n x n identity and zero matrices, respectively.

25.4 Inner Product Revisited The inner product was defined in Chapter 1 in terms of a metric function lbat took two vectors as inpnt and manufactored a number. We now know what kind of machine this is in lbe language of tensors. symmetric bilinear form

25.4.1. Definition. A symmetric bilinear form 9 on V is a symmetric tensor of type (0, 2). If {ej }r~l is a basis of V and {.i}~l is its dual basis, lben 9 = gij.i. j (recall Einstein'ssummation convention),because ei eJ = ei @ ej + e j ® ei forma basis of 82 (V). For any vector v E V, we can write 9(V)

= gij.i. j (v) = gij.i. j (vkekl = vkgij.i . j (ek) = gijV j•i. '-,,-'

=!if

(25.16)

Thus, 9(V) E V*. This shows that 9 can be thought of as a mapping, 9 : V --+ V*, given by Equation (25.16). For lbis equation to make sense, it should not matter which factor in lbe symmetric product v contracts wilb. But this is a trivial

750

25. ALGEBRA OF TENSORS

consequence of the symmetries gij = gji and f;.i f) = €i e', Thecomponents gijV i of g(v) in the basis lei}!:1 ofV' are denoted by Vi, so g(v) = Viei,

(25.17)

We have thus lowered the index of vj by the use of the synunetric bilinear form g. Iu applications Vi is uniquely defined; furthermore, there is a one-to-one correspondence between Vi and vi. This can happen if and only if the mapping 9 : V --+ V' is invertible, in which case there must exist a unique g-1 : V* --+ V, or g-I E S2(V*) = S2(V), such that vjej

= v = g-Ig(v) = g-I(Viei) = vig-I(e i) k

---

= Vi [(g_l/k ejek] (e') = Vi (g_l/ ej ek(e i) = Vi (g-l/, ej. =8~

Comparison of the LHS and the RHS yields vj = Vi (g-1 omit the -I and simply write

/i. It is customary to (25.18)

where it is understood that 9 with upper indices is the inverse of 9 (with lower indices). nondegenerate bilinear forms and inner products

25.4.2. Definition. An invertible bilinear form is called nondegenerate. A symmetric bilinear form that is nondegenerate is called an inner product.

raising and lowering indices

We therefore see that the presence of a nondegenerate symmetric bilinear form (or an inner product) naturally connects the vectors in V and V* in a unique way. Going from a vector in V to its unique image in V' is done by simply lowering the index using Equation (25.17), and going the other way involves using Equation (25.18) to raise the index. This process can be generalized to all tensors. For instance, although in general, there is no connection among 'J'5(V), 'J'!(V), and ':!1(V), the introduction of an inner product connects all these spaces in a natural way and establishes a one-to-one correspondence among them. For example, to a tensor in 'J'5(V) with components tij there corresponds a unique tensor in 'J'!(V), given, in component form, by tj = gjktik, and another unique tensor in ':!1(V), given by tij = giltj = gilgjkt1k. Let us apply this technique to gij, which is also a tensor and for which the lowering process is defined. We have (25.19)

This relation holds, of course, in all bases. The inner product has been defined as a nondegenerate synunetric bilinear form. The important criterion of nondegeneracy has eqnivalences:

25.4 INNER PRODUCT REVISITED

751

25.4.3. Proposition. A symmetric bilinear form 9 is nondegenerate if and only

if

1. the matrix ofcomponents gij has a nonvanishing determinant, or 2. for every nonzero v E V, there exists WE V such that g(v, w)

t= o.

Proof The first part is a direct consequence of the definition of nondegeneracy. The second part follows from the fact that 9 : V --> V' is invertible iff the nullity of 9 is zero. It follows that if v E V is nonzero, then g(v) t= 0, i.e., g(v) is not the zero functional. Thus, there must exist a vector w E V such that [g(v)](w) t= O. The proposition is proved once we note that [g(v)](w) sa g(v, w). 0 25.4.4. Definition. A general (not necessarily nondegenerate) symmetric bilinear form 9 can be categorized as follows:

1. positive (negative) definite: g(v, v) > 0 [g(v, v) < 0] for every nonzero

vector v; 2. definite: 9 is either positive definite or negative definite; 3. positive (negative) semidefinite: g(v, v) '" 0 [g(v, v)

:s 0] for every v;

4. semidefinite: 9 is either positive semidefinite or negative semidefinite; 5. indefinite: 9 is not definite.

25.4.5. Example. Some of the categories of the definition above can be illustrated in :[R2 with vi = (xj , Yd, v2 = (X2, Y2), and v = (x, y). (a) Positive definite: g(Vl, Vl) = XIX2 +YIY2 because if v 'f:. 0, then one of its components is nonzero, and g(v, v) x 2 + y2 > O. (b) Negative definite: g(vl. V2) = !(XjY2 + X2Y1) - XjX2 - YIY2 because g(v, v) = xy - x 2 - y2 = (x - y)2 _ x 2 y2. which is negative for nonzero v.

=

-!

!

!

(c) Indefinite: g(vl. V2) = XjX2 - YIY2. For v = (x, x), g(v, v) = O. However, 9 is nondegenerate, because it hasthe invertible matrix 9 = ~1) in the standard basis of~2. (d) Positive semidefinite: g(Vl. V2) = XIX2 ::::} g(v, v) = x 2 andg(v, v) is never negative.

(6

However, 9 is degenerate because its matrix in the standard basis oflR.2 is 9 = is not invertible.

g-orlhogonal and null vectors

(6 g), which II

Two vectors u, v E V are said to be g-orthogonal ifg(u, v) = O. A null vector of9 is a vector that is g -orthogonal to itself. If 9 is definite, then the only null vector is the zero vector. The converse is also true, as the following proposition shows. 25.4.6. Proposition. If 9 is not definite, then there exists a nonzero null vector.

Proof That 9 is not positive definite implies that there exists a nonzero vector v E V such that g(v, v) :s o. Similarly, that 9 is not negative definite implies that there exists a nonzero vector w E V such that g(w, w) '" O. Construct the vector u = av + (1 - a)w and note that g(u, u) is a continuous function of a. For a = 0 this function has the value g(w, w) '" O,andfora = lit has the value g(v, v):S O. Thus, there must be some a for which g(u, u) = O. 0

752

25. ALGEBRA OF TENSORS

25.4.7. Example. In the specialtheory of relativity, the inner prodoetof two "position" four-vectors, rt = (xj , Yl, z j , efl) and rz = (X2. Y2, Z2, et2), where c is the speed of light, is defined as

This is clearly an indefinite symmetric bilinear form. Proposition 25.4.6 tells us that there must exist a nonzero null vector. Such a vector r satisfies

g(r, r) = e2 t 2 -

x2 -

i - z2 = 0,

x 2 + y2 + z2 ,;x2 or c 2 = 2 ::::} c = ±

+ y2 + Z2 t

=

distance

±-.--. This corresponds to a

time particle moving with the speed of light. Thus, light rays are the null vectors in the special t

theoryof relativity.

III

Whenever there is an inner product on a vector space, there is the possibility of orthogonal basis vectors. However, since g(v, v) is allowed to be negative or zero, it is impossible to demand normality for some vectors. g-orthonormal vectors, diagonal components of g, index and signature a! 9

polarization identity

25.4.8. Definition. A basis (eil~t of V is g-orthonormal if g(ei, ej) = 0 for i i' i. and g(ei, e.) (nosumf) is + I, -I, orO. The g(ei, ei) are called the diagonal components ofg. We use n+, n_, and no to denote the number of vectors ei for which g(ei, ei) is, respectively, + I, -I, or O. The integer n.: is called the index of g, and s = n+ - n_ is called the signature of g. The existence of orthonormal bases was established for positive definite g by the Gram-Schmidt orthonormalization process in Chapter I. One of the steps in this process is division by (the square root 01) g(v, v), which could be zero in the present context. A slight modification of the Gram-Schmidt process, in conjunction with the procedure used in the symplectic case, can help construct an orthonormal basis for a general (possibly indefinite) g. To be specific, start with the polarization identity, g(v, v') = ![g(v + v', v + v') - g(v - v', v - v')], and use it to convince yourself that 9 is identically zero unless there exists a vector e) such that g(e), ej) i' O. Rescaling, we can assume that g(et, et) == ~) = ±1. Let V) be the span of et and V2 the g-orthogonal complement of Vt. Clearly, Vt n V2 = (OJ; moreover, VI + V2 = V, forifv E V, we can write v=

~tg(v,

eirer + v - mg(v, et)et .

~.

eV1

eV2

Now, if 9 = 0 on V2, we are done; otherwise, there is a vector 02 such that g(e2, e2) sa ~2 = ±1. We continue indnctively to complete the construction. We therefore have the following:

25.4 INNER PROOUCT REVISITED

Sylvester's theorem

753

25.4.9. Theorem. For every symmetric bilinearform g on V, there is an orthonormol basis. Furthermore; n+o n_, and no are the same in all orthonormal bases.

Orthonormal bases give the same oriented volume element.

For a proof of the last statement, also known as Sylvester's theorem, see [Bish 80, p. 104]. Orthonormal bases allow ns to speak of the oriented volume element. Suppose leilf=1 is an oriented basis of V*. If IcI'}f~1 is another orthonormal basis in the same orientation and related to lei} by a matrix R, then

eZ /\ ... /\ eN.

Since l
25.4.10. Example. Let V

= lll.3 and vi = (XI> YI, ZI), vz = (xz, YZ, zz), and v = (x, Y, z). Define the synunetric bilinear form 9(VI, vz) = !(X1YZ +xZYI

+ YIZZ +YZZI +xIZZ +XZZI)

sothat g(v, v) = xy-l-yz-l-xz, Wewishto find asetof vectorsin]g3 that areorthonormalwith respectlo Ibisbilinearfonn. Clearly, et = (I, 1,0) is such thatg(el, el) = 1. So ej is one of our vectors.Considerv = (1,0, I) and notethat the vectorrj = v- [g(v, el)/g(el, el)lel> suggested by the Gram-Schmidt process, is orthogonal to ej . Furthermore, it is easily Therefore, our second vectoris verifiedthatg(r2. r2) =

-i-

rZ

02 = "Ilg(rz, rz)1 =

(I -

3 2)

-"IS' - -"IS' -"IS

with g(eZ,02) = -1. Finally,we take w = (0, I, I). Then r3

=w -

g(w, el) ei _ g(w, 02) 02 = 4 (-3, I, I) g(el>el) g(02, ez) III

will be orthogonal to both ej and 02 with g(r3, r3) = -~. Thus, the !bird vector can be chosento be

r3 e3 = v'lg(r3, r3)1

(3 I I)

= - -"IS' -"IS' -"IS ,

°

and we obtain g(el, el) = I, g(eZ,02) = -I, g(e3' e3) = -I, g(e" ei) = for i i' iWealso have n+ = 1, n_ = 2, and no = 0, i.e., the index of 9 is 2, audits signature is -1. Although we haveworked in a particular basis, Theorem 25.4.9 guarantees that n+. n_. and no are(orthonormal) basis-independent. II We should emphasize that the invariance of n+, n_, and no is true for gorthonormal bases. As a counterexample, consider g of Example 25.4.10 applied to the standard basis of R3, which we designate with a prime. It is readily verified that g(e;, e;) = 0

for i = 1,2,3.

754

25. ALGEBRA OF TENSORS

So it might appear that no = 3 for this basis. However, the standard basis is not g-orthonormal. In fact, ' ') = '21 =g (el,e3 ' ') =g (e'Z,e3' ') 9 ( el,e Z

That is why the nonstandard vectors ei, v, and w were chosen in Example 25.4.10. In an orthonormal basis the matrix of 9 is diagonal, with entries +I, -I, and O. In particular, if 9 is to be nondegenerate, that is, to be an inner product, no must be zero. Thus, a general inner product on an N-dimensional vector space satisfies the conditions n+ + n_ = N and n+ - n_ = s which give s = N - 2n_.

25.4,11. Definition. An inner product space with n.: = I or n : = N - I is called a Minkowski space. For N = 4 this is the space ofthe special theory ofrelativity. An inner product space with n.: = 0 (or n_ = N) is called a Euclidean space.

25.4.12. Example. Let {eil~1 permutation tensor

be a basis of

V and {ei}~l its dual. We can define the

permutation tensor: ~ili2 .•.iN "'J' . it 2···JN

) =eit Aei2 A···AeiN( eJ'l,eJ'2,···,ejN·

(25.20)

It is clear from this definition that D)i,Il)'? ••i ,! is completely skew-symmetric in all upper 1 2...JN indices. That it is also skew-symmetric in the lower indices can be seen as follows. Assume that two of the lower indices are equal. This means baving two e j 's equal in (25.20). These

two e j 's will contract with two ei '5, say e k and el . Thus, in the expansion there will be a term C.!'(ej).f(ej), where C is the product of all the other factors. Since the product is completely skew-symmetric in the upper indices, there must also exist another term, with a minus sigo and in which the upper indices k and I are interchanged: -Cet(ej).f«ej)' This makes the sum zero, and by Theorem 25.2.7, (25.20) is antisymmetric in the lower indices as well. This suggests that ~i.li~ .. .i!f ex Eiliz ...iN E" ' . To find the proportionality constant }I1Z ... }N JI1Z .. ·}N ' we note that (see Problem 25.16) ,12 N " ,I ,2 ,N "12 N = L., E,,(I),,(2) ...,,(N)",,(I)",,(2) ... ",,(N)' zr

The only contribution to the sum comes from the permutation with the property n(O = i, This is the identity permutationfor which Err = 1. Thus, we have ~g:::~ = 1. On the other hand, e I2... N EI2 ...N = EI 2...N = (-1)"-. Therefore, the proportionality constant is (_l)n- . Thus Eitiz...iNE"

.

_ (_l)n_~iliz iN

}lJ2 .. ·}N -

hh ts

III

We can lind an explicit expression for the permutation tensor of Example 25.4.12. Expanding the RHS of Equation (25.20) using Equation (25.10), we obtain

(25.21)

25.4 INNER PRODUCT REVISITED

755

Furthermore, the first equation of (25.21) can be written concisely as a determinant. To see this, first note that ~ili2

"12

iN

N

~

I'!iz

i'ir

'il:iN

= L..., E"CI)"C2)..."CN)""CI)""C2) ... ""CN)· n

The RHS is clearly the determinant of a matrix (expanded with respect to the ith row) whose elements are The same holds true if 1,2, ... , N is replaced by h, n. ... .i», thus,

81'.

..

.

. . . -_ det

8lIl2 ... IN

JIl2···}N

s:;2

;1 f/?

["

;N

," J 8l~ IN

{/~ J2

it :

f/nN f/N J2

(25.22)

(/N,

IN

25.4.13. Example. Let us applythe secondequationof (25.21) to Euclidean lll.3: €

ijk

,i ,j ,k

,i ,j sk

,i ,j ,k

€lmn = UtUmUn - VI"nvm - UrnOt "n

+ UmUnUt ,i ,j sk + ,i ,j sk unul Urn -

,i ,j ,k

UnUmUt"

From this fundamental relation, we can obtain other useful formulas, For example, setting

n = k and summing over k, we get €

ijk

_ 3,i,j

€lmk -

,i,j

ulum - ulum -

3,i ,j UrnUt

+ UrnOt ,i ,j + ,i ,j UrnUt -

,i,j _ ,i,j

,i ,j

CJlf)m - ulum - umoZ .

Now set m = j in this equation and sum over j: cik ... ell fljk = 38i - 8i = 28i.

= i and sumover i: . €ijk = 28; = 2 . 3 = 31.

Finally, let 1 ··k

ell

In general, eitiz ...iNe··

.

- (_I)n-NI

IPZ···IN -

.,

eiliZ .. .iN-liNe· . . .' 1I 12··· IN_llN -

(-l)n-(N _l)ISiN

(/liz •.. iN_2 iN_liN E" , . . llI2··· lN_21N_IlN

• iN'

= (-l)n-(N _ 2)1 (f/N-18j~ _ 8i!i- 1 si,N ) IN-1 IN IN IN-l'

~OO~



25A.14. Example. As an application of the foregoing formalism, we can express the determinant ofa 2 x 2 matrix in tenus of traces. LetA be such a matrix with elements A~. Then detA = EijA\ A~

= !(EijA\ A~ :- EijAA{) = !(EijEkZ A~A{)

_lijkl

kl_l

ij

jj

- 2AkAZ(8i8j -8 j8i) - 2(A iA j -AjAi) = ![(trA)(lrA) - (A2)iJ = ![(lrA)2 _tr(A2)].

III

756

25. ALGEBRA OF TENSORS

We can generalize the result of the example above and express the determinant of an N x N matrix as

(25.23)

25.5

The Hodge Star Operator

It was established in Chapter 2 that all vector spaces of the same dimension are isomorphic (identical). Therefore, the two vector spaces AP(V) and AN-p(V) having the same dimension, (~) = (N~)' must be isomorphic. In fact, there is a natural isomorphism between the two spaces: Hodge star operator

25.5.1. Definition. Let 9 be an inner product and [e;}~1 an ordered gorthonormal basis of V. The Hodge star operator is a linear mapping, AP(V) --> AN-P(V), given by (remember Einstein's summation convention!)

*

* (Cit /\ ... A ei

p

ip+l .. .iN I ) es (N _ p)!f:i 1.. .ip ei p + l

/\ ... A eiN'

(25.24)

where all the indices of the < tensor are raised by gij. Although this definition is based on a choice of basis, it can be shown that the • . s: b" de d m th at Eip+l .. .iN ± Eit .. .ipip+t .. .iN be In ract asis-m pen ent. we note ... ;;;;;; operatorIS il

fp

cause each index raising introduces either a +lora -1. In particular, for Euclidean spaces, in which n_ = 0, the two epsilon symbols are the same.

25.5.2. Example. Let us appty Definition 25.5.t 10 AP(R3) for p let. e2. e3} be an oriented orthonormal basis ofR3. (a) For AO(R3) = JR a basis is I, aod (25.24)gives *1

= 0, 1,2,3. Lei

= -3!IE lj"k ej 1\ eJ' 1\ ek = ei 1\ ez A eg,

(b) For A I(R3) = JR3 a basis is {e" e2, e3), aod (25.24) gives *ei = *el = e2 1\ eg, 'sez = e3 A CI, *e3 = ei A e2·

-ir.(kej /\ ek- or

(e) For A2(JR3) a basis is [e1/\ e2, ej A e3, e2 /\ eg}, aod (25.24)gives .ei /\ ej =
The preceding example may suggest that applying the Hodge star operator twice (composition of * with itself, or * 0 *) is equivalent to applying the identity operator. This is partially troe. The following theorem is a precise statement of this conjectme. (For a proof, see [Bish 80, p. Ill].)

25.5 THE HOOGE STAR OPERATOR

757

25.5.3. Theorem. Let V be an oriented space with an inner product g. For A E AP(V), we have

* 0 *A '" * *A = (-Ij"-(-I)p(N-P)A,

(25.25)

where n.: is the index oig and N = dim V. In particular, for Euclidean spaces with an odd number of dimensions (such as 3 1R ) , A = A. One can extend the star operation to any A E A P (V) by writing A as a linear combination of basis vectors of AP(V) constructed out of {eil~I' and using the linearity of *. Io the discussion of exterior algebra one encounters sums of the form

**

antlsyrnmetnc tensors with numerical coefficients

Ait··.ipvII · A··· /\

v'I p '

It is important to keep in mind that Ail".i p is assumed skew-symmetric. For example, if A = ei /\ ez, then in the sumA = Aij ei A. ei- the nonzero components consist of A 12 = and A 21 = Similarly, when B = ei 1\ e2 1\ e3 is written in the form B = Bijk ei 1\ ej 1\ ek, it is understood that the nonzero components of B are not restricted to B 123. Other components, such as B 132, B 231, and so on, are

!

-!.

also nonzero. In fact,we have B 123

= _B132 = _B213 = B231 = B 312 = _B 321 = ~.

This should be kept in mind when sums over exterior products with numerical

coefficients areencountered. Crossproduct is defined onlyinthree dimensions!

25.5.4. Example. Let a, bE ]R3 and (eloe2, e3} an oriented orthonormal basis of]R3. Then a = aiel andb = biej. Letus calculate a A b and*(a Ab). Weassume a Euclidean 9 on 1R3. Thena A b = (aief) A (hiej) = albici 1\ ej. and *(a 1\ b)

= *(aiei) 1\ (,) ej) = a i,) * (ei 1\ ej) = ai,) (Etek) =

(Efja i ,))ek.

Weseethat «(aAb) is avectorwithcomponents [*(3Ab)]k = Etaibi, whichareprecisely thecomponents of ax b. Thecorrespondence between a A b anda x b holds only in three dimensions, because dbnA 1(\7) = dbn A2(\7) only if dim \7 = 3. Thatis why tbe crossproduct can be definedas a "machine" that takes two vectors in V andmanufactures a vector in V-only if V is three-dimensional. II

25.5.5. Example. We can use tbe resnlts of Examples 25.4.13 and 25.5.4 to establish a sample of familiar vector identities componentwise. (a) Forthetriple crossproduct, we have [a x (b x c)]k

= Etai (b x c)j = Etai (4mbl em) = aibl c"l Ekij EjIm

=ai bi emE kij EImj = ai b1c"'(,k,i 01 om -

,koi)

0mOI

= aibkci - aibick = (a' c)bk - (a- b)ck,

758

25, ALGEBRA OF TENSORS

which is the kth componentof b(a . c) - c(a ' b). In deriving the above "bac cab" rule, we used the factthatone can swapanupper indexwith the samelowerindex: a i hi = ai bi (b) Next we show the familiar statement that the divergence of curl is zero.Let 8i denote differentiation withrespect to Xi. Then V. (V x a)

= ai(V

x ali

= aif~kajak =

'~

F _El' Bj8jak

.~

= -El!

= -a/V x a)j =

fijk8i8jak

8j8jak

'~

= -8j(E j l

- V . (V x a)

=}

BiOk)

V, (V x a) = 0,

The abovesteps show in general that

25.5.6.Box. When the product of two tensors is summed over a pair of indices in which one of the tensorsis symmetricand the other antisymmetric, the result is zero.

(c) Finally, we showthat curlof gradient is zero: [V x (V f)]k

because Eijk



= f~k8j8k f = fijk8jakf = 0,

is antisymmetric in jk, while 8j8kf is symmetric in jk.

III

25.6 Problems 25.1. Show that the mapping v : V* --+

~

given by v( T) = T(V) is linear.

25.2. Show that the components of a tensor product are the products of the components of the factors:

25.3. Show that eh 181

181 ej, 181 €il 181 ... 181 €i, are linearly independent. Hint:

Consider A J,,') ..,j, eJ') 181 181 eJ', 181 €il 181 .. ' 181 €i, = 0 and evaluate the LHS on 1···l s appropriate tensors to show that all coefficients are zero.

25.4. What is the tensor product of A = 2ex - e y 25.5. If A show that

E

+ 3ez with itself?

.G(V) is represented by A~ in the basis {eil and by A;k in {ep. then

A'ke' iO\ e'l 101 ""j 1 k'O' - Aij e.I'O'~' where {€j} and {e'l} are dual to Ie;} and {ep, respectively.

25.6. Prove that the linear functional F : V --+ ~ is a linear invariant. i.e., basisindependent, function.

25.6 PROBLEMS

25.7. Show that tr :

759

'Ji --> ]R is ao invariaot linear function.

25.8. If A is skew-symmetric in some pair of variables, show that S(A) = O. 25.9. Using the exterior prodnct show whether the following three vectors are linearly dependent or independent: VI = 2el'- e2 +3e3 - e4, V2 = -ret +

3~

- 2e4,

V3 = 3el + 2e2 - 4e3 + e4. 25.10. Let A E 'J(j(V) be skew-symmetric. Show that if r l , ... ,r' E V* are linearly dependent, then A(r l , ... , T'") = O. 25.11. Show that (ek A e.} with k < i are linearly independent. 25.12. Let v E Vbenonzero, aod let A E AP(V). Show that v AA = 0 ifaodonly if there exists B E A p-I (V) such that A = v A B. Hint: Let v be the first vector of a basis; separate out v in the expaosion of A in terms of the p-fold wedge products of basis vectors, aod multiply the resnlt by v. 25.13. Let A E A 2(V) with components A ij . Show that A A A = 0 if.and only if Aii A k1 _ A ik Ai i + Ail Ai k = 0 for all i, j, k, I in aoy basis. 25.14. A linear operator acting on a vector in one dimension simply multiplies that vector by some constaot. Show that this constaot is independent of the vector chosen. That is, the constaot is ao intrinsic property of the operator. 25.15. Let {el, e2, e3} be aoy basis in ]R3. Define ao operator E : ]R3 --> ]R3 that permutes aoy set of three vectors {VI,V2, V3} to {Vi, vi, Vk}. Find the matrix representation of this operator aod show that det E = 'ijk. 25.16. (a) Starting with the definition of the permutation tensor 8Jil'J·2 ...iJN , aod I 2·" N writing the wedge product in terms of the aotisynnnetrized tensor product, show that

(b) Show that the inverse of the (diagonal) matrix of gin ao orthonormal basis is the same as the matrix of g. (c) Now show that .12...N = (-I)'-'12...N = (-1)'-. 25.17. Let{edf:" be a g-orthonormal basis ofV. Let ~ii =

±8} be the matrix of

9 in this orthonormal basis. Let {Vi Jf=1 be aoother (not necessarily orthonormal) basis of V with a traosformation matrix R. Using G to denote the matrix of 9 in {Vi}, show that det G = det 7](det R)2 = (-1)'- (det R)2.

760

25. ALGEBRA OF TENSORS

In particular, the sign of this determinant is invariant. Why is det G not equal to det 'I)? Is there any conflict with the statement that the determinant is basisindependent? 25.18. Show that the kemel of 9 : V --> V* consists of all vectors u E V such that g(u, v) = 0 for all v E V. Show also that in the g-orthonormal basis lei J, the set lei Ig(ei, ei) = OJ is a basis of kerg, and therefore no is the nullity of g. 25.19. Use Equation (25.23) to show that for a 3 x 3 matrix A,

25.20. Show that a 2-form w is nondegenerate if and only if the determinant of (wii) is nonzero if and only if wb is an isomorphism. 25.21. Let V be a finite-dimensional vector space and w E A 2 (V*). Suppose there exist a pair of vectors ej , ei E V such that w(e\, ei) i' O. Let p\ be the plane spanned by e\ and ei, and V\ the w-orthogonal complement of P i- Show that V\ n p\ = 0, and that v - w(v, eiJe\ + w(v, e\)ei is in VI. 25.22. Show that Lj~\ .1 1\ .i+n , in which l.iJ7=\ is dual to led!",\, the canonical basis of V, has the same matrix as w. 25.23. Suppose that v, Vi cients {Xi, Yi, Zi} and {x;,

E

V are expressed in a canonical basis of V 'with coeffi-

y;, z~}. Show that

n

w(v,

Vi)

= ~)XiY[ - X[Yi). i=l

25.24. Let V be a vector space and V* its dual. Define w w(v

+
where v, space.

Vi E

V and
E

E


V*. Show that (V Ell V*, w) is a symplectic vector

25.25. By taking successive powers of w show that n

wk =

L

A 2 (V Ell V*) by

e it 1\ e h +n 1\ ... /\ e ik /\ e A+n.

it ... jk=l

Conclude that

where [n/2] is the largest integer less than or equal to n/2.

25.6 PROBLEMS

761

25.26. Show that the condition for a matrix A to be symplectic is AtJA = J where J = (~I A) is the representation of w in the canonical basis. 25.27. Show that Sp(V, w) is a snbgroup of GL(V). 25.28. Find the index and the signature for the bilinear form 9 on ~3 given by g(VI, V2) = XI)'2 +X2YI - YIZ2 - Y2ZI·

25.29. In relativistic electromagnetic theory the current J and the electromagnetic field tensorF are, respectively, a four-vector" and an antisymmetric tensor of rank 2. Thatis,J = Jk ek and F = Fijei /\ej. Find the components ohJ and *F. Recall that the space of relativity is a 4D Minkowski space. 25.30. Show that where there is a sum over an upper index and a lower index, swapping the upper index to a lower index, and vice versa, does not change the sum. In other words, Ai B, = Ai B i . 25.31. Show the following vector identities, using the definition of cross products

in terms of Eijk. (a) A x A =0.

x B) = (V x A)· B - (V x B)· A. (e) V x (A x B) = (B . V)A + A(V . B) - (A· V)B - B(V . A). (b) V· (A

(d) V x (V x A) = V(V· A) - V 2A.

25,32. A vector operator Y is defined as {yl, y2, y3}, a set of three operators, satisfying the following commutation relations with angular momentum: [yi, Jj] = i.ijkYk. Show that ykYk commutes with all components of angular momentum. 25.33. The Pauli spin matrices I

(0 I)

0"=10' describe a particle with spin these matrices satisfy

0"

3=

(I0-10)

t in nonrelativistic quantum mechanics. Verify that

[ui,u i ] =uiu J --'-uJa i =2iE~ak,

{ui,uJ}=uia J +aJa i =2B~12'

where 12 is the unit 2 x 2 matrix. Show also that O"i O"j = i.~ O"k any two vectors a and b, (0"" a)(O"'· b) = a- b12 + iO"'· (a x b).

+ 8} 12, and for

25.34. Show that any contravariant tensor ofrank two can be written as the sum of a symmetric tensor and an antisymmetric tensor. Can this be generalized to tensors of arbitrary rank? 6It turns out to be more natural to consider J as a 3-form.However,such a fine distinctionis not of any consequencefor the presentdiscussion.

762

25. ALGEBRA OF TENSORS

Additional Reading 1. Abraham, R., Marsden, J. and Ratiu, T. Manifolds, Tensor Analysis, and Applications, 2nd ed., Springer-Verlag, 1988. A comprehensive textbook on tensors with many examples drawn from physics. 2. Bishop, R. and Goldberg, S. Tensor Analysis on Manifolds, Dover, 1980. This little and lucid book is one of the earliest ones on index-free tensor analysis. 3. Flanders, H. Differential Forms with Applications to Physical Sciences, Dover, 1989. One of the first books on exterior algebra written for physicists. It has many examples drawn from various areas of physics.

26,

----'-

_

Analysis of Tensors

Tensor algebra deals with lifeless vectors and tensors--objects that do not move, do not change, possess no dynamics. Whenever there is a need for tensors in physics, there is also a need to know the way these tensors change with position and time. Tensors that depend on position and time are called tensor fields and are the snbject of this chapter. In stndying the algebra of tensors, we leamed that they are generalizations of vectors. Once we have a vector space V and its dual space V*, we can take the tensor products of factors of V and V* and create tensors of various kinds. Thus,

once we knowwhata vectoris, we canmakeup tensors from it. In theprevious chapter, we did not concern ourselves with whata vectorwas; we simply assumed that it existed. Because all the vectors considered there were stationary, their mere existence was enough. However, in tensor analysis, where things keep changing from point to point (and over time), the existence of vectors at one point does not guarantee their existence at all points. Therefore, we now have to demand more from vectors than their mere existence. Tied to the concept of vectors is the notion of space, or space-time. Let us consider this first.

26.1 Differentiable Manifolds Space is one of the undefinables in elementary physics. Length and time intervals are concepts that are "God given," and any definitions of these concepts will be circular. This is true as long as we are confined within a single space. In classical physics, this space is the three-dimensional Euclidean space in which every motion takes place. In special relativity, space is changed to Minkowski space-time. In nonrelativistic quantum mechanics, the underlying space is the (infinite-dimensional) Hilbert space, and time is the only dynamical parameter. In the general theory

764

26. ANALYSIS OF TENSORS

of relativity, gravitation and space-time are intertwined through the concept of curvature. Mathematicians have invented a unifying theme that brings the common features of all spaces together. This unifying theme is the theory of differentiable manifolds. A rigorous understanding of differentiable manifolds is beyond the scope of this book. However, a working koowledge of manifold theory is surprisingly simple. Let us begin with a crude definition of a differentiable manifold. differentiable manifold provisionaily defined

26.1.1. Definition. A differentiable manifold is a collection of objects called points that are connected to each other in a smoothfashion such that the neighborhood of each point looks like the neighborhood ofan m-dimensional (Cartesian) space; m is called the dimension ofthe manifold.

As is customary in the literature, we use "manifold" to mean "differentiable manifold." 26,1.2. Example. The following are examplesof differentiable manifolds. (a) The space R" is an a-dimensional manifold. (b) The surface of a sphere is a two-dimensional manifold. (c) A torus is a two-dimensional manifold. (d) The collection of all n x n real matrices whose elements are real functions having derivatives of all orders is an n 2-dimensional manifold. Here a point is an n x n matrix.

(e) The collectionof all rotations in 1R3 is a three-dimensional manifold. (Here a point is a rotation.) (f) Any smooth surface in]R3 is a two-dimensional manifold. (g) The unit n-sphere Sn, which is the collection ofpoints in lRn+ I satisfying Xl2

+---+Xn2+ l = I ,

is a manifold. Any surface with sharp kinks, edges, or points cannot be a manifold. Thus, neither a cone nor a finite cylinder is a two-dimensional manifold. However, an infinitely long III cylinder is a manifold.

coordinate functions and charts

eOO-related charts and atlases

Let Up denote a neighborhood of P. When we say that this neighborhood looks like an m-dimensional Cartesian space, we mean that there exists a bijective map rp : Up -> IRm from a neighborhood Up of P 10 a neighborhood rp(Up) of rp(P) in IRm , such that as we move the point P continuously in Up, its image moves continuously in rp(Up). Since rp(P) E IRm, we can define functions xi : Up -> IR such that rp(P) = (x1(P), x 2(P), ... , xm(P». These functions are called coordinate functions of tp, The numbers xi (P) are called coordinates of P. The neighborhood Up together with its mapping rp fomn a chart, denoted by (Up,rp). Now let (Vp, 1-') be another chart at P with coordinate functions I-'(P) = (yl(p), y2(p), ... ,ym(p» (see Figure 26.1). It is assumed that the function I-' 0 rp-l : rp(Up n V p) -> I-'(Up n Vp), which maps a subsel of IRm to another subset of IRm, possesses derivatives of all orders. Then, we say that the two charts

26.1 DIFFERENTIABLE MANIFOLDS

765

_ - - - - -_ _ M

_-+

Jlf.m

Figure 26.1 1\vo charts (Up, 'fJ) and (Vp, JL), containing P are mapped into Ill.m. The function JL 0 cp -1 is an ordinary functionfrom IRm to JRm.

f-L and 'fJ are oo -related. Such a relation underlies the concept of smoothness iu the definition of a manifold. A collection of charts that cover the manifold and of which each pair is oo-related is called a oo atlas.

e

e

e

26.1.3. Example. For the two-dimensional unit sphere S2 we can construct an atlas as follows. Let P

= (x, y, z) be a point in S2. Then x 2 + y2 + z2 = I, or

Theplus sign corresponds to the upperhemisphere, andthe minus sign to the lowerhemisphere. Let be the upper hemisphere with the equator removed. Then a chart 'fJ3) with 'fJ3 : ]R2 can be constructed by projecting on the xy-plane: 'fJ3(P) = (x, y). Similarly, (U;, JL3) with JL3 : U; ..... ]R2 given by JL3(P) = (x, y) is a chart for the lower hemisphere. In manifold theory the neighborhoods on which mappings of charts are defined have no boundaries (thus the word "open"). Thisis becauseit is moreconvenient to define limits on boundaryless (open) neighborhoods. Thus, in the above two charts the equator, which is the boundary for bothhemispheres, mustbe excluded.Withthis exclusion and Ui cannot cover the entire S2; hence, they do not form an atlas. More charts are needed to coverthe unittwo-sphere. 1\\'0 suchcharts arethe rightandleft hemispheres and for which y > 0 and y < O. respectively. However, these two neighborhoods leave two points uncovered, the points (1,0,0) and (-I, 0, 0). Again this is because boundaries of the right and left hemispheres must be excluded. Adding the frontand backhemispheres to the collection coversthese two points.Then S2 is completelycoveredandwe have

ut ut . . .

Construction ofan atlas forthe sphere S'.

(ut,

Ui

ut

ut

U2',

766

26. ANALYSIS OF TENSORS

an atlas.There is, of course,a lot of overlapamong charts.Wenow showthat these overlaps are eoo-related. As an illustration, we consider the overlap between and This is the upper-right

ut Ui.

quarter of the sphere. Let

(Ui, 1'3) aud (Ui, 'fJ2) be charts with

qJ3(x, y, z) = (x, y),

qJ2(X, y, z) = (x, z}.

The inverses are therefore given by 1'3 1(x, y)

= (x,

y, z)

= (x,

y, Jt - x 2 - y2),

1',1 (x, z) = (x, y, z) = (x, ~t - x 2 - z2, z), aud 1'20 1'3 1 (x, y) = qJ2(x, y, JI-x 2 - y2) = (x, JI-x2 _ y2). Let us denote CfJ2 0 components of F:

qJ3 1 by

Fl(X, y) = x

aud

F, so that F : JR2 --+ ]R2 is descnbed by two functions, the

The first component has derivatives of all orders at all points. The second component has derivatives of all orders at all points except at x 2 + y2 = 1, which is excluded from the region of overlap of and for which z can never be zero. Thus, F has derivatives of all orders at all points of its domain of definition. One can similarly show that all regions of overlap for all charts have this property, i.e., all charts are oo-related. IIIlI

ut vt,

e

26.1.4. Example. For 82 of the preceding example, we cau find a new atlas in terms

of new coordinate functions. Since X[ + xI + x~ = I, we can use spherical coordinates e = cos- 1 x3, I' = tau-I(X2/Xl). A chart is theu given by (8 2 - {I} - (-I), J1.), where

Jk(P) = (8, rp) maps a point of 8 2 onto a region in IR2. This is 'schematically shown in Figure 26.2. The singletons {I} aud {-I} are the north aud the south poles, respectively. This chart cannot cover all of 8 2, however, because when 0 (or :n'), the value of the azimuthal angle rp is not determined. In other words, = 0 (or :n') determines one point

e

e=

of the sphere (the north pole or the south pole), but its image in IR2 is the whole range of q; values. Therefore, we must exclude e = 0 (or rr) from the chart (8 2, Jk). To cover these two points, we need more charts. IIIlI

illustration of stereographic projection

26.1.5. Example. A third atlas for 82 is the so-called stereographic projection shown in Figure 26.3. In such a mapping the image of a point is obtained by drawing a line from the north pole to that point and extending it, if necessary, until it intersects the Xl xz-plane. It cau be verified that the mapping I' : 82 - {I} -+ ]R2 is given by I' (x to X2, x3) =

(~,~). I-X3

1- X3

We see that this mapping fails for X3 = I, that is, the north pole. Therefore, the north pole must be excluded (thus, the domain 8 2 - {Ij). To cover the north pole we need another stereographic projection-this time from the south pole. Then the two mappings will cover all of 8 2, and it can be shown that the two charts are eoo-related (see Example ~I.I~ •

,

26.1 DIFFERENTIABLE MANIFOLDS

767

Figure 26.2 f)

= 0,

1C,

A chart mapping points of S2 into }R2. Note that the map is not defined for and therefore at least one more chart is required to cover the whole sphere.

The three foregoing examples illustrate the following fact, which can he showu to hold rigorously:

26.1.6. Box. It is impossible to cover the whole S2 with just one chart.

vector spaces are manifolds

product manifold defined

26.1.7. Example. Let V be an m-dirnensional real vector space. Fix any basis {es I in V with dual basis [e'}. Define : V -e- ~"' by (v) = (el(v), ... , e"'(v». Then the reader may verify that ('\7, tP) is an atlas. Linearity of ¢ ensures that it has derivatives of all orders. This construction shows that V is a manifold of dimension m. 11II

If M and N are manifolds ofdimensions m and n, respectively, we can construct their product manifold M x N, a manifold of dimension m + n. A typical chart on M x N is obtained from charts on M and N as.follows, Let (U, ",) be a chart on M and (V, ",) one on N. Then a char1 on M x N is (U x V, '" x ",) where '" x ",(P,

submanifold

open submanifolds

Q) = (",(P), ",(Q» E jR"' x jRn =

jR",+n

for P E U, Q E V.

26.1.8. Defiuition. Let M be a manifold. A subset N ofM is calleda submanifold of M if N is a manifold in its own right. A trivial, but important, example of submanifolds is the so-called open submanifold. If M is a manifold and U is an open subset! of M, then U inherits a manifold structme frnm M by taking any chart (U., "'.) and restricting "'. to U n Ua- It is clear that dim U = dim M. Having gained familiarity with manifolds, it is now appropriate to consider maps between them that are compatible with their structme.

IRecall that an open subset U is one each of whose points is the center of an open ball lying entirely in U.

768

26. ANALYSIS OF TENSORS

Figure 26.3 Stereographic projection of S2 into R 2. Note that the north pote has no image under this map; another chart is needed to cover the whole sphere.

differentiable maps and their coordinate expressions

function as a special kind of map

26.1.9. Definition. Let M and N be manifolds ofdimensions m and n, respectively. Let f : M ~ N be a map. We say that f is eoo , or differentiable, iffor every chart (U,
diffeomorphism and local diffeomorphism defined

26.1.11. Definition. A bijective differentiable map whose inverse is also differentiable is called a diffeomorphism. Two manifolds between which a diffeomorphism exists are called diffeomorphic. Let M and N be manifolds. M is said to be diffeomorphic to N at P E M if there is a neighborhood U of P and a diffeomorphism f: U ~ feU). Then f is called a local diffeomorphism at P.

2The domain of J.L 0 f 0 (,0-1 is not all of R", but only its open subset rp(U). However, we shall continue to abuse the notation and write jRm instead of rp(U). This way, we do not have to constantly change the domain as U changes. The domain is always clew from the context. Although Proposition 16.4.6 was shown for Donned linear spaces, it really holds for all "spaces" for which the concept of open set is defined.

769

26.1 DIFFERENTIABLE MANIFOLDS

diffeomorphlsms of a manifold form a group n-sphere and its stereographic projection

In our discussion of groups, we saw that the set of linear isomorphisms of a vector space Vanta itself forms a group GL(V). The set of diffeomorphisms of a manifold M onto itself also forms a group, which is denoted by Diff(M). 26.1.12. Example. Thegeneralization of a sphere is theunit a-sphere, which is a subset of JRn+ I defined by Sn

= (xI, .. "

Xn+l) E JRn+ll xf

+ ... +X~+l = I).

The stereographic projection definesan atlasfor Sn as follows. For all pointsof Sn except (0,0, ... , I), the north pate, define the cbart 9'+ : S" - (II '" U+ -+ JRn by 'P+(Xb ...• Xn+ l ) = (

X l , ...• t

1 - xn+ l

x

n

.)

- xn+ l

Toinclude thenorth pole, consider a secondchart ip.: by CP-(xI •... , Xn+ l ) = (

Xl

1 +X n+l

, ... , I

X

n

+X n+l

for(xl. ...• Xn+l)EU+. : Sn -

{-I}

==

U- -+ R" defined

)

Next,let us find theinverses of thesemaps. Wefind theinverse of 'P+: that of C}J- can be found similarly. Let ~k ee Xk/(t - x"H)' Then one can readily sbow that

f--.2 I+xn+t L5k =

=>

I - Xn+t

k=l

xn+ l

=

EZ-l~l+1

EZ=1 ~k2 -

I

and 2~i

for i

Xi = -:-1-_-=,,=n=--=."2

=

1,2, ... , n.

L..k=1 >k

From thedefinition of CP+. we have 9':;:1 (~j, ... , ~n) = (XI, ... , Xn, Xn+l) = (

. 2?

2"'"

I - Ek = l ~k

(26.1)

2~

2'

E~-l~~+I).

I - Ek = l ~k E k= l ~k - I

On theoverlap of U+ and U-, i.e., on all points of Sn exceptthenorth and thesouth poles,
on a typical entry of Eqnation (26.1): 2~j

Therefore,

-I

.)


(~I n

2"'"

E k= l ~k

)' n~n 2 E k= l ~k

1

It is clear that CfJ- 0 CfJ'+ has derivatives of all orders exceptpossiblyat a pointfor which ';i = 0 forall i. But this would correspond toxn+ l = 1, which is excluded from the region

of overlap.

II

770

26. ANALYSIS OF TENSORS

M

f

IFIlm

Figure 26.4 Corresponding to every map f JLO f o
26.2

:M

---+ N there exists a coordinate map

Curves and Tangent Vectors

We noted above that functions are special cases of Definition 26.1.9. Another special case occurs when M = R. This is important enough to warrant a separate definition. differentiable curve

26.2.1. Definition. A differentiable curve in the manifold M is a eoo map of an

interval of~ to M.

initial and linal points 01 a curve

This definition should be familiar from calculus, where M = ~3 and a curve is given by its parametric equation (11 (t), h(t), /3(t)), or simply by ret). The poiot yea) E M is called the initial point, and y(b) E M is called the final point of the curve y. A curve is closed if yea) = y(b). We are now ready to consider what a vector at a poiot is. All the familiar vectors io classical physics, such as displacement, velocity, momentum, and so forth, are based on the displacement vector. Let us see how we can generalize such a vector so that it is compatible with the concept of a manifold. In ~2, we define the displacement vector from P to Q as a directed straight lioe that starts at P and ends at Q. Furthermore, the direction of the vector remaios the same if we connect P to any other final poiot on the lioe P Q located beyond Q. This is because ~2 is a flat space, a straight lioe is well-defined, and there is no ambiguity io the direction of the vector from P to Q. Thiogs change, however, if we move to a two-dimensional spherical surface such as the globe. How do we define the straight lioe from New York to Beijiog? There is no satisfactory definition of the word "straight" on a curved surface. Let us say that "straight" means shortest distance. Then our shortest path would lie on a great circle passiog through New York and Beijiog. Define the "direction" of the trip as the "straight"

26.2 CURVES AND TANGENT VECTORS

771

arrow, say I km in length, connecting our present position to the next point I km away. As we move from New York to Beijing, going westward, the tip of the arrow keeps changing direction. Its direction in New York is slightly different from its direction in Chicago. In San Francisco the direction is changed even more, and by the time we reach Beijing, the tip of the arrow will be almost opposite to its original direction. The reason for such a changing arrow is, of course, the curvature of the manic fold. We can minimize this curvature effect if we do not go too far from New York. If we stay close to New York, the surface of the earth appears flat, and we can draw arrows between points. The closer the two points, the better the approximation to flatness. Clearly, the concept of a vector is a local concept, and the process of constructing a vectoris a limiting process. The limiting process in the globe example entailed the notions of "closeness." Such a notion requires the concept of distance, which is natural for a globe but not necessary for a general manifold. For most manifolds it is possible to define a metric that gives the "distance" between two points of the manifold. However, the concept of a vector is too general to require such an elaborate stmcture as a metric. The abstract usefulness of a metric is a result of its real-valucdness: given two points PI and P2, the distance between them, d(Pj, P2), is a nonnegative real number. Thus, distances between different points can be compared. We have already defined two concepts for manifolds (more basic than the concept of a metric) that together can replace the concept of a metric in defining a vector as a limit. These are the concepts of (real-valued) functions and curves. Let us see how functions andcurves can replacemetries. Let y : [a, b] --> M be a curve in the manifold M. Let P E M be a point of M that lies on y such that y(c) = P for some c E [a, b]. Let f E FOO(P). Restrict f to the neighboring points of P that lie on y. Then the composite function f 0 y : R --> R is a real-valued function on R. We can compare values of f 0 y for various real numbers close to c-as in calculus. If U E [a, b] denotes" the variable, then f 0 y(u) = f(y(u)) gives the value of f 0 y at various u's. In particular, the difference !'1(f 0 y) sa f(y(u» f(y(c)) is a measure of how close the point y(u) E M is to P. Going one step further, we define

d(~: y) I.=c = Y-Tc f(Y(U)~

=

:(y(C)) ,

(26.2)

the usual derivative of an ordinary function of one variable. However, this derivative depends on y and on the point P. The function f is merely a test function. We could choose any other function to test how things change with movement along y. What is important is not which function we choose, but how the curve y causes it to change with movement along y away from P. This change is determined by 4We usuallyuse u or t to denotethe (real)argument of the map y : [a, b] --+ M.

772

26. ANALYSIS OF TENSORS

the directional derivative along y at P, as given by (26.2). A directional derivative determines a tangent which, in turn, suggests a tangent vector. That is why the tangent vector at P along y is defined to be the directional derivative itself! The use of derivative as tangent vector may appear strange to the novice, especially physicists encountering it for the first time, but it has been familiar to mathematicians for a long time. It is hard for the beginner to imagine vectors being charged with the responsibility ofmeasuring the rate of change of functions. It takes some mental adjustment to get used to this idea. The following simple illustration may help with establishing the vector-derivative connection. illustration ofthe equality ofvectors and directional derivatives

26.2.2. Example. Let us take the familiar case of a plane and consider the vector a = axe x + ayeyo What kindof a directional derivative can correspond to a? First we need a curve y : IR -+ ]R2 that is somehow associated with a. It is not hard to convince oneself thatthemost natural association is that of vectorsto tangents. Thus,we seek a curvewhose tangent is (paral1el to) a. The easiest (but not the only) way is simply to take the straight line alonga; that is, let y(u) = (axu, ayu). Thedirectional derivative att = 0 for anarbitrary function f : ]R2 --+ lR is given by d(f

y)

0

I

du

u=O

= lim f(y(u» - f(y(O» = lim f(ax u, ayu) - f(O, 0). U

U--70

u-*O

U

(26.3)

Taylor expansion in two dimensions yields

f(ax u, ayu) = f(O, 0)

afl + aeu -a

afl + ayu -a

x u=o

y

u=o

+ ....

Substituting in (26.3), we obtain

d(f 0 y) I du

u=Q

= lim axu(af/ax)u~o u-+O

af = ax ax

+ ayu(af/ay)u~o + ... U

af + ay ay

(a a) = ax ax + ay ay f.

Thisclearly showstheconnection between directional derivatives andvectors. In fact, the correspondences aIax # ex and alay # ey establish this connection verynaturally. Note that thecurve y chosenabove is by no means unique. In fact,there areinfinitely many curves that have the same tangent at t = a andgive the same directional derivati~.



Sincevectors arethesameasderivatives, we expectthemtohavetheproperties shared by derivatives: tangent vector defined

26.2.3. Definition. Let M be a differentiable manifold. A tangent vector at P

E M is an operator t : Foo(P) --+ R such that for every f, g E Foo(P) and ex, fJ E R

1. t is linear: t(exf + fJg) = ext(!) derivation properly of tangent vectors

+ fJt(g);

2. t satisfies the derivation property: l(fg) = g(P)t(f)

+ f(P)I(g).

26.2 CURVES AND TANGENT VECTORS

tangent space defined

773

The operator! is an abstraction ofthe derivative operator. Note thatt(f), g (P), f(P), and t(g) are all real numbers. The reader may easily check that if addition and scalar multiplication of tangent vectors are defined in an obvious way, the set of all tangent vectors at P E M becomes a vector space, called the tangent space at P and denoted by 'Jp (M). If U is an open subset of M (therefore, an open submanifold of M), then it is clear that 'Jp(U) = 'Jp(M)

foraliPEU.

(26.4)

Definition 26.2.3 was motivated by Equations (26.2) and (26.3). Let us go backwards and see if (26.2) is indeed a tangent, that is, if it satisfies the two conditions of Definition 26.2.3. vectors tangentto a curve

e

26.2.4. Proposition. Let y be a oo curve in M such that y(c) = P. Define ;Y(c) by

~ d (J(c))(f) sa du f

0

I

Y u~c

for every f E FOO(P). Then;Y(c) is a tangent vector at P called the vector tangent to y ate. Proof We have to show that the two conditions of Definition 26.2.3 are satisfied for f, g E FOO(P) and 01, fJ E R The first condition is trivial. For the second condition, we use the product rule for ordinary differentiation as follows:

~ d d (J(c))(fg) = du (fg) 0 i Y u~c ;;: du [(f = [:u (f

0

y)luJ (g

= [(;y(c»(f)]g(y(c)) = [(;Y(c»(f)]g(P)

0

0

y)(g

0

y)] Iu~c

Y)u~c + (f Y)u~c [:u (g 0

0

y)luJ

+ f(y(c))[(;Y(c))(g)]

+ f(P)[(;Y(c))(g)].

Note that in going from the first equality to the second, we used the fact that by definition, the product of two functions evaluated at a point is the product of the values of the two functions at that point. D Let us now consider a special curve and corresponding tangent vector that is of extreme importance in applications. Let


e

774

26. ANALYSIS OF TENSORS

coordinate curve, coordinate vector field, and coordinate frames

26.2.5. Definition. Let (Up, rp)beachartat P EM. Thenthe curve y' : lR -+ M, defined by yi (u) = rp-I (xl (P), ... , xi-I(p), u, xi+l(p), ... , xm(P)) is called the ith coordinate curve through P. The tangent vector to this curve at P is denoted by ai Ip and is called the ith coordinate vector field at P. The collection ofall vector fields at P is called a coordinate frame at P. The variable u is arbitrary in the sense that it can be replaced by any (good) junction ofu.

Let c = xi (P). Then for f

E

F""(P), we have

• d (ailp)f = (Yi(C»(f) = dJ =

0

y il u=c

.!:.- f(rp-I(xl(p), ... ,xi-l(p), u,xi+l(p), ... ,xm(pml du

== ::i I p

=}

ailp =

a~i Ip ,

u=c

(26.5)

where the last equality is a (natural) definition of the partial derivative of f with respect to the ith coordinate evaluated at the point P. This partial derivative is again a e"" function at P. We thereforehave the following: 26.2.6. Proposition. The coordinate frame (ai IpIr=1 at P is a set of operators ai(P) : F""(P) -+ lR given by (ailp)f= afl

ax'

(26.6) p

== dd f(rp-I(xl(p), ... , xi-l(p), u, xi+l(p), ... , xm(p»)1 U

. u=c

Another conunon notation for af/ax i is f.i 26.2.7. Example. Pick a point p = (sinB cosrp, sinB sinrp, cosB) onthesphere S2 in a chart (Up, 1") given hyI"(sin Bcosip, sinB sinip, cosB) = (B, rp).1f B is kept constant and qJ

is allowedto varyovervaluesgiven by u, thenthe coordinate curveassociated with q; is

given by y~(u) = I"-I(B, u) = (sinB cosu, sinB sinu, cosB).

As u varies, ytp (u) describes a curve on S2. This curve is simply a circle of radius sin e. The tangent to thiscurve at anypointis fJ/ocp, or simply arp, thederivative with respect to

thecoordinate ip, Similarly, thecorve Ye (u) describes agreat circle on S2 with tangent ae ee a/ aB.

III

The vector space T p (M) of all tangents at P was mentioned earlier. In the case of 8 2 this tangent space is simply a plane tangent to the sphere at a point. Also, the two vectors, ae and a~ enconnteredin Example26.2.7 are clearlylinearly independent.Thus, they form a basis for the tangent plane. This argumentcan be generalizedto any manifold. The following theorem is such a generalization(for a proof, see [Hish80, pp. 51-53]):

775

26.2 CURVES AND TANGENT VECTORS

26.2.8.Theorem. Let M be an m-dimensional manifold and P Remember Einstein's

E M. Then the set {aiIp}7::dorms a basis oj'J'p(M). In particular, 'J'p(M) is m-dimensional. An arbitrary vector, I E 'J'p (M), can be written as

summation convention!

Thelast statement can be derived by letting both sides operate on xi and usiog Equation (26.6). Let M = V, a vector space. Choose a basis {e.} io V with its dual considered as coordinate functions. Then, at every v E V, there is a natural isomorphism e : V --> 'J'v(V)mappiogavectoru = aiei E Vontoaiailv E 'J'v(V). The reader may verify that this isomorphism is coordioate iodependent; i.e., if one chooses any other basis of V with its corresponding dual, then .p(v) will be the same vector as before, expressed in the new coordinate basis. Thus, 26.2.9. Box. IjV is a vector space, thenjor all v

E

V, one can identify 'J'v (V)

with V itself.

Suppose we have two coordinate systems at P, {xi} with tangents ailp and {yi} with tangents Vi Ip. Any IE 'J'p(M) can be expressed either io terms of ailp or in terms of Vi Ip: I = a i ailp = piVi Ip. We can use this relation to obtaio a i in terms of pi: From Theorem 26.2.8, we have

ai

= I(x i) = (piVilp}(xi) ==

[pi

~I ](xi) =pi ay} p

In particular, if t = Vklp, then pi = I(yi) = [Vklp](yi) a i = ax i /a/. Thus, usiog Equation (26.5),

aXi.J . ay} p

(26.7)

= 81, and (26.7) gives (26.8)

For any function j

E

FOO(P), Equation (26.8) yields

This is the chaio rule for differentiation.

26.2.10. Example. Let us find the coordinate curves and the coordinate frame at

P = (x, y, z) on S2. Weuse thecoordinates of Example 26.1.3. In particular, consider ({J3. whose

inverse is givenby

776

26. ANALYSIS OF TENSORS

The coordinate corve y.!(u) along y is obtainedby letting y be a function5 of u:

where h(O) = y and h'(O) f E FOO(P), and note that

= .'!...- f(Y.!(u)) I

azf

du

= a,

u=o

a constant. To find the coordinate vector field at P, let

= .'!...- f(x, h(u),

[1

du

JI- x 2 - h2(u)) Iu=O

I + -af -(-2h(u)) dh 1 ] ay du u=o az 2 du ,}I- x 2 - h 2(u) u=o f =a(a _?'.-af)=a(~_?'.-~)f. ay z az ay z az = -af -dh

So, choosing the function h in such a way that a = 1,

az = ay-

y

-ar, z

a

where ay and z are the coordinate vector fields of IR3. The coordinate vector field be obtainadsimilarly.

al can ill

26.3 Differential of a Map Now that we have constructed tangent spaces and defined bases for them, we are ready to consider the notion of the differential (derivative) of a map between manifolds.

differential ofa map at a point

26.3.1. Definition. Let M and N be manifolds ofdimensions m and n, respectively, and let 1ft : M --> N be a eoo TYUlp. Let P E M, and let Q = 1ft(P) E N be the image of P. Then there is induced a map 1ft,P : 'Jp(M) --> 'JQ(N), called the differential of1ft at P and given asfollows. Let IE 'J p(M) and f E FOO(Q). The action of1ft,p(l) E 'J Q(N) on f is defined as (1ft,p(I»(f)

== l(f 0

1ft).

(26.9)

The reader may check that the differential of a composite map is the composite of the corresponding differentials, i.e., (1ft

0


(26.10)

Furthermore, if 1ft is a local diffeomorphism at P, theu 1ft,p is a vector space isomorphism. The inverse of this statement-which is called the inverse mapping theorem, and is much harder to prove (see [Abra 88, pp. 116 and 196]}--is also true:

26.3 DIFFERENTIAL DF A MAP

inverse mapping theorem

777

26.3.2. Theorem. (inverse mapping theorem) If 1ft : M --+ N is a map and 1ftoP : 'Jp(M) --+ 'Jy,(p)(N) is a vector space isomorphism, then 1ft is a local dijfeomOlphism at P.

Let us seehow Equation(26.9)looks in temnsof coordinatefunctions.Suppose that {xili'=l are coordinates at P and {yal:=l are coordinates at Q = 1ft(P). We note that ya o 1ft is a real-valued eoo function on M. Thus, we may write (with the function expressedin temns of coordinates)

We also have 1= a i ailp. Similarly, 1ftop(l) = tJa(a/aya)IQ because {a/aya)IQl fomn a basis. Theorem 26.2.8 and Definition 26.3.I now give tJa = 1ftop(I)(ya) = I(ya

0

1ft) = l(fa)

at" m at" =[aiailp](t")=ai-.I ="'ai_·I· ax' p L., ax' p 1=1

This can be written in matrix fomn as

l tJtJ2 )

· ( ·· tJn

m) l) m 2 afl/aX (a af2/ax a

(afl/aXI af2/ax l

.. .

·· ·

af"/ax l

afn/ax m

..

.'

(26.II)

am

Jacobian matrix of of The n x m matrix is denoted by J and is called the Jacobian matrix of 1ft with a differentiable map respect to the coordinates xi and ya. On numerous occasions the two manifolds are simply Cartesianspaces, so that 1ft : ~m --+ R", In such a case, f" is naturally written as 1ft", and the Jacobian matrix will have elements of the fomn a1ft"/ax i. Differential a An important special case of the differential of a map is that of a constant constant map isthe map. Let 1ft : M --+ {Ql E N be such a map; it maps all points of M onto a single zero map. point Q of N. For any f E Foo(Q), the function f 0 1ft E FOO(P) is constant for all P EM. Let I E 'J p (M) be an arbitrary vector. Then (1ftop(I))(f) es l(f

0

1ft) = 0

(26.12)

because I(c) = 0 for any constant c. So,

26.3.3.Box. If 1ft : M --+ {Q lEN is a constant map, so that it maps the entire manifold M onto a point Q of N, then 1ftoP : 'Jp(M) --+ 'J Q(N) is the zero map,

SSeethe laststatement of Definition 26.2.5.

778

26. ANALYSIS OF TENSORS

1\\'0 other special cases merit closer attention: M = JR for arbitrary N, and N = JR for arbitrary M. In either case 'J,(JR) is one-dimensional with the basis vector (dldu) I,. When M = JR, the mapping becomes a curve, y : JR --> N. The only vector whose image we are interested in is 1= (dldu)I" with y(c) = P. From (26.9) using Proposition 26.2.4 in the last step, we have

[ yo, ddU Ic ] f

=

dd f U

0

yI

u=c

= (f(c»(f).

This tells us that the differential of a curve at c is simply its tangent vector at y (c). It is common to leave out the constant vector (dldu) I" and write y*, for the LHS. components of tangenfs to curves

26.3.4. Example. Itis usefulto haveanexpression forthecomponents of thetangent to a curve y atan arbitrary pointon it. Since y mapsthe real line to M. with a coordinate patch established on M. we can write y as y = (yl, ... , ym) where r' = xi 0 yare ordinary functions of one variable. Proposition 26.2.4 then yields y*t!

d d f(y j (u), ... , y ml = -du f 0 yi d = -du f(y(u)) Iu=t = -du (u» u=1 u=t = af

dyi ax' du

I u=1

es

af dyi = ax' dt

·yiad,

or (26.13)

Forthisreason, Y*t is sometimes denoted by y.

differentialofa real-valued function

When N = JR, we are dealing with a real-valued function f : M --> R The differential of fat P is fop : 'Jp(M) --> 'J,(IR), where c = f(P). Since 'J,(JR) is one-dimensional, for a tangent IE 'Jp(M), we have f*p(l) = a(dldu)I,. Let g : JR --> JR be an arbitrary function on R Then [f*p(l)](g) = a(dgldu)" or, by definition of the LHS, l(f 0 g) = a(dgldu),. To find a we choose the function g(u) = u,i.e., theidentityfunctionjthen zigyds = I andl(f og) = l(f) = a. We thus obtain f.p(t) = l(f)(dldwlk. Since 'J,(JR) is a flat one-dimensional vector space, all vectors are the same.and there is no need to write (dldu)I,. Thus, we define the differential of f, denoted by df sa f., as a map df : 'Jp(M) --> JR given by

df(l) = l(f).

(26.14)

In particnlar, if f is the coordinate function xi and I is the tangent to the jth coordinate curve aj Ip, we obtain (26.15)

26.3 DIFFERENTIAL DF A MAP

779

This shows that

26.3.5. Box. {dx' [p

Remember Einstein's summation convention!

}?::I is dual to the basis {ajlp }j=1 of'Jp(M).

26.3.6. Example. Letf: M -+ Rbe areal-valued functionnn M. Let xi becnnrdinates at P. Wewant to express df in terms of coordinate functions. Fort E 'Jp (M) we can write t = a i ailp and df(t)

= t(f) = ar'[8;1p ](f) = ,,!a,(f),

wbere in the last step, we suppressed the P. Theorem 26.2.8 and Equation (26.14) yield

ee' = t(x') = (dx')(t). We thus bave df(t)

= 8,(f)[(dx')(t)] = [8i(f)(dx i)](t).

Sincethisis true for all t, we get . ~ . ~8f . df = 8,(f)(dx') sa L.. 8, ( f) (dx ' ) = L.. -,dx'. i=l

i=l

(26.16)

ax

This is theclassical formula for the differential of a function f. If we choose s', the jth member of a newcoordinate system, for f. we obtain

(26.17) which is the transformation dual to Eqnation(26.8).

III

The following is a powerful theorem that constructs a submanifold out of a differeutiable map (for a proof, see [Warn 83, p. 31]): 26.3.7. Theorem. Assume than? : M -+ N is a e"" map, that Q is a point in the range of 1/J, and that 1/J. : 'Jp(M) --'> 'J Q(N) is surjective for all P E 1/J-I(Q). Then 1/J-I(Q) is a submanifold of M and dim 1/J-I(Q) = dimM - dimN. Compare this theorem with Propositiou 26.1.10. There, V was an open subset of N, and since (V) is open, it is automatically an open submanifold. The difficulty in proving Theorem 26.3. 7 lies in the fact that 1/J-I (Q) is closed because {Q}, a single point of N, is closed. We can justify the last statement of the theorem as follows. From Equation (26.12), we readily conclude that 'Jp (1/J-I (Q)) = ker 1/J.p. The dimension theorem, applied to 1/J.P : 'Jp(M) --'> 'J Q(N), now gives

r:'

dim 'Jp(M)

= dimker1/J.p + rank1/J.p

=}

dimM

= dim 1/J-I(Q) -+- dimN,

where the last equality follows from the surjectivity of 1/J.p.

780

26. ANALYSIS OF TENSORS

26.3.8. Example. Consider a eoo map / : R" -+ lEI. Let c E lEI snch that the partial derivatives of / aredefined andnotallzeroforall pointsof /-1 (c).Then,according toEqnation (26.11), a vector cia, E Tp(lEIn ) is mappedhy /. to the vector (Jl' (aJ/ax') f=cd/dt. Since ai/ax' are not all zero,hyproperly choosing(Jl', we can make (Jl' (a//ax') f=cd/dt sweep overall real numbers. Therefore, f* is surjective, andby Theorem 26.3.7, j-ICc) is an (n - I)-dimensional submanifold of R". A noteworthy special case is the function defined by

/(x!, x2 , ... , xn ) =

(x l ) 2 + (x2 )2 + ... + (xn )2

andc = r 2 > O. Then,/-I (c), an (n-I)-sphereofradins r, is a submanifold ofR".

III

26.4 Tensor Fields on Manifolds So far we have studied vector spaces, learned how to construct tensors ont of vectors, tonched on manifolds (the abstraction of spaces), seen how to construct vectors at a single point in a manifold by the use of the tangent-at-a-curve idea, and even found the dual vectors dx' Ip to the coordinate vectors 8, Ip at a point P of a martifold. We have everything we need to study the analysis of tensors.

26.4.1 Vector Fields We are familiar with the concept of a vector field in 3D: Electric field, magnetic field, gravitational field, velocity field, and so forth are all familiar notions. We now want to generalize the concept so that it is applicable to a general martifold. To begin with, let us consider the following definition.

26.4.1.Definition. The union ofall tangent spaces at different points 0/a manifold tangent bundle defined

M is denoted by T(M) and called the tangent bundle of M: T(M) =

U 'Jp(M) PeM

It can be shown ([Bish 80, pp. 158-164]) that T (M) is a manifold ofdimension 2dimM. vector field defined

vector fields related bya map

26.4.2.Definition. A vector field X on a subset U ofa manifold M is a mapping X : U -+ T(M) such that X(P) es Xlp sa X» E 'Jp(M). The set of vector fields on M is denoted by X(M). Let M and N be manifolds and F : M -+ N a differentiable map. We say that the two vector fields X E X(M) and Y E X(N) are F -related if F.(X p) = Y F(p)/or all P EM. This is sometimes written simply as F.X = Y. It is worthwhile to point out that F.X is not, in general, a vector field on N. To be a vector field, F.X mnst be defined at all points of N. The natural way to define F.X at Q E N is [F.X(Q)](f) = X(f 0 F) where P is the preimage of Q, i.e.,

26.4 TENSOR FIELDS ON MANIFOLDS

781

F(P) = Q. But there may not exist any such P (F may not be onto), or there may be more than one P (F may not be one-to-one) with such property. Therefore, this natural construction does not lead to a vector field on N. If F.X happens to be a vector field on N, then it is clearly F -related to X. In terms of the coordinates xi, at each point P E M,

X p '" Xlp = X~Bilp, where the real numbers X~ are components of X» in the basis {Bilp[.As P moves around in U, the real numbers X~ keep chimging. Thus, we can think of X~ as a function of P and define the real-valued function Xi : M --+ R by Xi (P) ea X~. Therefore, the components of a vector field are real-valued functions on M. 26.4.3. Example. Let M = ]R3. At each point P = (x, y, z) E ]R3, let (ex, ey, e z) be a basis for }R3. Let Vp be the vector space at P. Then TOR 3 ) is the collection of all vector spaces V p for all P. We can determine the value of an electric field at a point in JR3 by first specifying the point, as Po = (xo, YO, zo), for example. This uniquely determines the tangentspace 'JPo(lR.3). Once we have the vector space, we can ask what the components of the electric field are in that space. These components are given by three numbers: Ex(xQ. YO, ZO), Ey(xO, YO, zo), and Ez(xo, YO, ZO)· The argoment is the samefor any othervectorfield. To specify a "point" in TOR 3 ) , we need three numbers to determine the location in }R3 and another three numbers to determine the components of a vector field at that point. Thus, a "point" in TOR3 ) is given by six "coordinates" (x, y, z, Ex. E y, E z), and T(}R3) is a six-dimensional manifold.

III

We know how a tangent vector t at a point P E M acts on a function f E FOO(P) to give areal numbert(f). We can extend this, point by point, for a vector field X and define a function X(f) by [X(f)](P) '" Xp(f),

P

E

U,

(26.18)

where U is a subset of M on which both X and f are defined. The RHS is welldefined because we know how Xp, the vector at P, acts on functions at P to give the real number [Xp ](f). On the LHS, we have X(f), which maps the point P onto a real number. Thus, X(f) is indeed a real-valued function on M. We can therefore define vector fields directly as operators on oo functions satisfying

e

+ fJg) = otX(f) + fJX(g), X(fg) = [X(f)]g + [X(g)]f.

X(otf

e oo vector fields

A prototypical vector field is the coordinate vector field Bi • In general, X(f) is not a eoo function even if f is. A vector field that produces a eoo function X(f) for every eoo function f is called a eoo vector field. Such a vector field has components that are eoo functions on M. The set of tangent vectors 'J p (M) at a point P E M form an m-dimensional vector space. The set of vector fields X(M)-which yield a vector at every point

782

26. ANALYSIS OF TENSORS

of the manifold-also constitutes a vector space. However, this vectur space is (uncountably) infinite-dimensional. A property of X(M) that is absent in 'Jp (M) is composition." This suggests the possibility of defining a "product" ou X(M) to tum it into an algebra. Let X and Y be vector fields. For X 0 Y to be a vector field, it has to satisfy the derivatiou property. But

x 0 Y(fg) = X(Y(fg)) = X(Y(f)g + fY(g)) = (X(Y(f)))g

+ Y(f)X(g) + X(f)Y(g) + f(X(Y(g)))

'I (X 0 Y(f))g + f(X 0 The set ofvector fields form a Lie algebra.

Y(g)).

However, the reader may verify that Xo Y - Y oX does iudeed satisfy the derivatiou property. Therefore, by defining the binary operation X(M) x X(M) --> X(M) as [X, Y]

== X 0

Y- Y

0

X,

X(M) becomes an algebra, called the Lie algebra of vector fields of M. The binary operation is called the Lie bracket. Although it was not mentioned at the time, we have encountered another example of a Lie algebra in Chapter 2, namely £, (V) under the binary operation ofthe commutation relation. Lie brackets have the following two properties: [X, Y] = -[V, X], [[X, V], Z]

Jacobi identity

+ [[Z, X], Y] + [[Y, Z], X] =

O.

These two relations are the defining properties of all Lie algebras. The last relation is called the Jacobi identity. X(M) with Lie brackets is an example of an infinite-dimensional Lie algebra; £, (V) with commutators is an example of a finite-dimensional Lie algebra. We shall have occasion to use the following theorem in our treatment of Lie groups and algebras in the next chapter: 26.4.4. Theorem. Let M and N be manifolds and F : M --> N a differentiable map. Assume that Xi E X(M) is F-related to Yi E X(N) for i = 1,2. Then [XI, X2] is F -related to [YI, Y2], i.e., F.[XI, X2] = [F.XI, F.X2].

Proof. Let

f

be an arbitrary fuoction on N. Then

(F. [XI , X2]) f es [Xl, X2](f

0

F) = XI (X2(f 0 F)) - X2 (XI(f 0 F))

= XI ([F.X2(f)]

0

F) - X2 ([F.XI(f)]

0

F)

= F.XI (F.X2(f)) - F.X2 (F.XI (f)) = [F.XI, F.X2] f, 6Recall that a typical element ofTp(M) is a map t : FOO(P) --+- R for which composition is meaningless.

26.4 TENSOR FIELDS ON MANIFOLDS

783

where we used Equation (26.9) in the first, second, and third lines, and the result of Problem 26.8 in the second line. D It is convenient to visualize vector fields as streamlines. In fact, most of the terminology used in three-diroensional vector analysis, such as flux, divergence, and curi, have their origins in the flow of fluids and the associated velocity vector fields. The streamlines are obtained-in nonturbulent flow-by starting at one point and drawing a curve whose tangent at all points is the velocity vector field. For a smooth flow this curve is unique. There is an exact analogy in manifold theory. integral curve of a vector field

26.4.5. Definition. Let X E X(M) be defined on an open subset U of M. An integral curve of X is a curve y whose range lies in U and for every t in the domain ofy, the vector tangent to y satisfies Yot = X(y(t)). Ify(O) = P, we say that y starts at P. Let us choose a coordinate system on M. Then X = Xi ili, where Xi are eoo functions on M, and, by 26.13, Y. = yi 8i. The equation for the integral curve of X will therefore become yi8i

= X i(y(t))8i,

or

dJri

= Xi (yl(t), ... , ym(t)),

i = 1,2, .. . ,m.

Since yi are siroply coordinates of points on M, we rewrite the equation above as i

dx i 1 m = X (x (t), ... , x (t)), dt

-

i = 1,2, ... , m.

(26.19)

This is a system of first-order differential equations that has a unique (local) solution once the initial value y(O) of the curve, i.e., the coordinates of the starting point P, is given. The precise statement for existence and uniqueness of iutegral curves is contained in the following theorem.

e

26.4.6. Theorem. Let X be a oo vector field defined on an open subset U of M. Suppose P E U, and c E R Then there is a positive number E and a unique integral curve y of X defined on It - c] :s E such that y(c) = P. 26.4.7. Example. EXAMPLES OF INTEGRAL CURVES (a) Let M = R with coordinate functionx. The vectorfield X = xax has an integral curve with initial point XD givenby the DE dxt dt = x(t), which has the solution x(t) = etxD. (b) Lei M = JR" with coordinate functions xi . The vectorfieldX = L a i Bi has an integral curve, with initial pointrD. givenby thesystemafDEs dx i / dt = a i • whichhasthesolution xi (r) = a i t + xb. or r = at + tu- Thecurve is therefore a straight line parallel to a going through rD.

(c) Let M X=

=

R" withcoordinate functions xi. Consider the vectorfield

L:" a}x i ai· i,j=l

784

26. ANALYSIS OF TENSORS

The integral curve of this vector field, with initial point ro- is given by the system of DEs dx i fdt = LJ=l a}x i , which can be written in vector fonn as dr tdt = Ar where A is a constant matrix. By differentiating this equation several times, one can convince oneself that dkr fdt k = Akr.The Taylorexpansionof ret) then yields

(d) Let M = ]R2 with coordinate x, y. The reader may verify that the vector field X = -yax + Xay has an integralcurvethrongh (xo, YO) givenby

x=xocost-Ya sint, y=xosint+YOcost, i.e., a circle centered at the origin passing through (xQ. Yo).

flow ofa vector field one-parameter group oftransformations

Going back to the velocity vector field analogy, we can think of integral curves as the path of particles flowing with the fluid. If we think of the entire fluid as a manifold M, the flow of particles can be thought of as a transformation of M. To be precise, let M be an arbitrary manifold, and X E :reM). At each point P of M, there is a unique local integral curve r» of X starting at P defined on an open subset U of M. The map F, : U --+ M defined by F,(P) = yp(t) is a (local) transformation of M. The collection of such maps with different t's is called the flow of the vector field X. The uniqueness of the integral curve yp implies that F, is a local diffeomorphism. In fact, the collection of maps (F, Item: forms a (local) one-parameter group of transformations in the sense that Ft

Global1-parameler group of transformations;

complete vector fields

III

0

F, = Ft +s ,

Fo = id,

(26.20)

One has to keep in mind that F, at a point P E M is, in general, defined only locally in t, i.e., only for t in some open interval that depends on P. For some special, but important, cases this interval can be taken to be the entire R for all P, in which case we speak of a global one-parameter group of transformations, and X is called a complete vector field on M. The symbol F , used for the flow of the vector field X does not contain its conuection to X. Iu order to make this conuection, it is cornmon to defiue F, sa exp(tX).

(26.21)

This definition, with no significance attached to "exp" at this point, converts Equation (26.20) into exp(tX)

0

exp(sX) = exp[(t

+ s)Xj,

exp(OX) = id, [exp(tX)]-1 = exp(-tX),

(26.22)

26.4 TENSOR FIELDS ON MANIFOLDS

785

which notationally justifies the use of "exp," We shall see in oor discussion of Lie groups that this choice of notation is not accidental.

Usingthisnotation, we canwrite Xp(f) '"

~f 0 F: (P)J = ~f dt ' ,=0 dt

One usually leaves out the function

x,

=

f

0

exp(tX)I

,=0

.

and writes

~ exp(tX) I ' dt

(26.23)

'=0

where it is understood that the LHS acts on some f that must compose on the RHS to the left of the exponential. Similarly, we have (F.X)F(P) G.F(P) \

(~F(exPtX)J ) dt ,=0

I

=

~F(exPtX)1 ' dt '=0

=

~G dt

0

F(exptx)1

,~O

' (26.24)

=F.(X)

where F : M --> N and G : N --> K are maps hetween manifolds. 26.4.8. Example. In this example, we derivea useful formula that gives the value of a functionat a neighboring point of P E M locatedon the integralcorve of X E X(M) going through P. We first note that since Xp is taugeutto yp at P = y(O), by Proposition26.2.4 we have

Nextwe use thedefinition of derivative and thefactthat Fo{P) = P to write . I hm - [f (F,(P)) - f(P)] = Xp(f). t

t-+O

Now,if we assume that t is verysmall, we have f (F,(P))

= f(P) + tXp(f) + ... ,

(26.25) III

which is a Taylor series withonly the first two terms kept.

26.4.2 Tensor Fields We have defined vector spaces 'J'p (M) at each point of M. We have also constructed coordinate bases, {adp Ii'=!, for these vector spaces. At the end of Section 26.2, we showed that the differentials {dx i Ip Ii'=! form a hasis that is dual to {ad p Ii':,,!. Let us concentrate on this dual space, which we will denote hy (M). Takiug the union of all 'JP(M) at all points of M, we obtain the cotangent

'JP

colangent bundle ofa manifold

786

26. ANALYSIS OFTENSORS

bundle of M:

T*(M) =

U 'Jj,(M).

(26.26)

PEM

This is the dual space of T (M) at each point of M. We can now define the analogue of the vector field for the cotangent bundle. differential one-term

pullback ofa differentiable map

26.4.9. Definition. A differential one-form 0 on a subset U ofa manifold M is a mapping 0 : U ~ T*(M) such that O(P) sa Op E 'Jj,(M). The collection ofall one-forms on M is denoted by X*(M).

If 0 is a one-form and X is a vector field on M, then O(X) is a real-valued function on M defined naturally by [O(X)](P) sa (Op)(X p). The first factor on the RHS is a linear functional at P, and the second factor is a vector at P. So, the pairing of the two factors produces a real number. A prototypical one-form is the coordinate differential, dx i . Associated with a differentiable map lfr : M ~ N, we defined a differential lfr* that mapped a tangent space of M to a tangent space of N. The dual of lfr* (Definition 1.3.17) is denoted by lfr* and is called the pullback of lfr. It takes a oneform on N to a one-form on M. In complete analogy to the case of vector fields, can be written in terms of the basis {dx i }: 0 = Oidxi. Here Oi, the components of 0, are real-valued functions on M. With the vector spaces 'J'p (M) and 'Jj,(M) at our disposal, we can construct various kinds of tensors at each point P. The union ofall these tensors is a manifold, and a tensor field can be defined as usual. Thus, we have the following definition.

o

bundle 01 tensors andtensorlieids

26.4.10. Definition. Let 'J'p (M) and 'Jj,(M) be the tangent and cotangent spaces at P E M. Then the set oftensors oftype (r, s) on 'J'p(M) is denoted by T;,p(M). The bundle oftensors oftype (r, s) aver M, denoted by T;(M), is

T;(M) =

U 'r.,p(M). PEM

A tensorfield T oftype (r, s) over a subset U of M is a mapping T : U such thatT(P) sa Tp sa Tip E T;,p(M).

~

T; (M)

10 particular, Tg(M) is the set of real-valued functions onM, TJ(M) = T(M), and T?(M) = T*(M). Furthermore, since T is a multilinear map, the parentheses are normally reserved for vectors and their duals, and as indicated in Definition 26.4.10, the value ofT at P E M is written as Tp or Tip. The reader may check that the map

T: X*(M) x .:. x X*(M), x X(M) x ... x X(M), ~ Tg(M) r times

s times

defined by

[T(w!, ... , w', VI, ... , v,)](P) = Tp(w!lp, ... , w'lp, vdp, ... , v,lp)

26.4 TENSOR FIELDS ON MANIFOLDS

Acrucial properly of tensors

787

has the property that

T( T(

+ gOi, , fVk + gUk,

, fw i

) = fT(

,wi,

) = fT(

, Vb

) + gT( )

+ gT(

, o', , Uk,

), )

(26.27) for any two functions f and g on M. Thus,7

26.4.11. Box. A tensor is linear in vectorfields and i-forms, even when the coefficients of linear expansion are functions.

The components of T with respect to coordinates xi are the m r +, real-valued functions

difficulty associated with differentiating tensors

If tensor fields are to be of any use, we must be able to differentiate them. We shall consider three types of derivatives with different applications. We stody one of them here, another in the next section, and the third in Chapter 2~. Derivatives can be defined only for objects that can be added (really, subtracted). For functions of a single (real or complex) variable, this is done almost subconsciously: We take the difference between the values of the function at two nearby points and divide by the length of the interval between the two points. We extended this definition to operators in Chapter 2 with practically no change. For functions of more than one variable, one chooses a direction (a vector) and considers change in the function along that direction. This leads to the concept of directional derivative, or partial derivative when the vector happens to be along one of the axes. In all the above cases, the objects being differentiated reside in the same space: f(t) and f(t + !'>t) are both real (complex) numbers; H(t) and H(t + At) both belong to 'c(V). When we try to define derivatives of tensor fields, however, we run immediately into trouble: T p and T p' cannot be compared because they belong to two different spaces, one to 'Y,.P (M) and the other to (M). To make comparisons, we need first to establish a "connection" between the two spaces. This connection has to be a vector space isomorphism so that there is one and only one vector in the second space that is to be compared with a given vector in the first space. The problem is that there are infinitely many isomorphisms between any given two vector spaces. No "natura!" isomorphism exists between 'J'rs, p (M) and 'J"s, p' (M); thus the diversity of tensor "derivatives!" We narrow down this diversity

'Y"p,

7In mathematical jargon, X(M) and X* (M) are called modules over the (ring of) real-valued functions on M. Rings are a generalization of the real numbers (field of real numbers) whose elements have all the properties of a field except that they may have no inverse. A module over a field is a vector space.

788

26. ANALYSIS OF TENSORS

by choosing a specific vector at T,'. p (M) and seeking a natnral way of defining the derivative along tihat vector by associating a "natural" isomorphism corresponding to tihe vector. There are a few methods of doingtihis. We describe one of tihemhere. First, let us see what happens to tensor fields under a diffeomorphism of M onto itself. Let F : M --+ M be such a diffeomorphism. The differential F.p of this diffeomorphism is an isomorphism of'J"p(M) and 'J"p(p)(M). This isomorphism induces an isomorphism of tihe vector spaces T;,p (M) and T;,p(p) (M)-also denoted by F.p-by Equation (25.7). Let us denote by F. a map of T(M) onto T(M) whose restriction to 'J"p(M) is F.p. 1fT is a tensor field on M, tihen F.(T) is also a tensor field, whose value at F(Q) is obtained by letting F.Q act on T(Q):

[F.(T)](F(Q)) = F.Q(T(Q)), or, letting P

= F(Q) or Q = F- 1(P),

[F.(T)](P) = F. P_l(P)(T(F- 1(P))).

(26.28)

Now, let X he a vector field and P E M. The flow of X at P defines a local diffeomorphism F, : U --+ F,(U) witih P E U. The differential F,. of tihis diffeomorphism is an isomorphism of'J"p (M) and 'J"P,(P) (M). As discussed above, this isomorphism induces an isomorphism of tihevector space T; (M) onto itself. The derivative we are after is defined by comparing a tensor field evaluated at P witihtihe image of tihe sarne tensor field under tiheisomorphism F,-; 1. The following definition makes this procednre more precise. Lie derivative of tensor fields with respect of avector field

26.4.12. Definition. Let P E M, X E X(M), and F, the flow of X defined in a neighborhood of P. The Lie derivative of a tensor field T at P with respect to X is denoted by (Lx T) p and defined by (26.29) Let us calculate tihe derivative in Equation (26.29) at an arbitrary value of t. For tihis purpose, let Q sa F,(P). Then d dt F,-;tTp,(P) ==

I

l~o /';t [F(-;-~tH). TpHO'(P) - F,-;lTp,(p)]

I = F,-;t lim A [Fi,~ Tp,+o,(P) - Tp,(p)] dt"""*O

ilt

1 = F,-;! lim - [Fi,~TFo,(Q) -TQ] .6..t~O 8t

== F,-;l (LxT)Q .

Since Q is arbitrary, we can remove it from tihe equation and write, as tihe generalization of Equation (26.29), LxT =

d -t F,.-F,. T. dt

(26.30)

26.4 TENSOR FIELDS ON MANIFOLDS

789

An important special case of the definition above is the Lie derivative of a vector field with respect to another. Let X, Y E :reM). To evaluate the RHS of (26.29), we apply the first term in the brackets to an arbitrary function f,

[F,-;;l Y F,(P)](f)

= Y F,(P) (f F;-l) = Y (f F,-l)IF,(p) = Y (f 0 L,)IF,(P) = Y (f - tX(f))IF,(P) 0

0

= (Y f) F,(P) - t Y(X(f))IF,(P)

+ t[X(Yf)]p - t{[Y(Xf)]p + t[X(Y(Xf))]p} = Yp(f) + tXp 0 Yp(f) - tYp 0 Xp(f) = Yp(f) + t [Xp, Yp](f) = Yp(f)+t [X, Y]p (f). = (Yf)p

The first equality on the first line follows from (26.9), the second equality from the meaning ofY F,(P); the second equality au the second line and the fourth line follow from (26.25). Finally, the fifth line follows if we ignore the t 2 term. Therefore,

(LxY)p (f) = lim! [F,-;;l y F,(p) - Yp] (f) t--*O t . 1 = lim - It [X, Y]p} (f) = [X, Y]p (f). t--+O t Lie derivative is commutator

Siuce this is true for all P and f, we get

LxY= [X, Y].

(26.31)

This and other properties of the Lie derivative are summarized in the following proposition.

Properties of Lie derivative

26.4.13. Proposition. Let T E T; (M) and T' be arbitrary tensor fields and X a given vector field. Then

1. Lx satisfies a derivation property in the algebra of tensors, i.e., LX(T@ T) = (LxT) @ T' + T@ (LxT').

2. Lx is type-preserving, i.e., LxT is a tensor field oftype (r, s). 3. Lx commutes with the operation ofcontraction oftensorfields; inparticular, in combination with property 1, we have Lx (9, Y) = (Lx9, Y)

+ (9, Lx.Y)

4. Lxf = Xf for every function f· 5. LxY = [X, Y]for every vector fieldY.

790

26. ANALYSIS OF TENSORS

Proof Except for the last property, which we demonstrated above, the rest follow directly from definitions and simple manipulations. The details are left as exercises.

o

Remember Einstein's summation

convention!

Although the Lie derivative of a vector field is nicely given in terms of commutators, no such simple relation exists for the Lie derivative of a I-form. However, if we workin a givencoordinate frame, then ausefulexpression fortheLie derivative of a I-form can be obtained. Applying Lx to (II, X), we obtain

Lx (II, Y) '----,.-'

= (LxII, Y) + (II, LxY) = (LxII, Y) + (II, [X, V]) . (26.32)

=X«O,y})

In particular, if Y = a, and we write X = xjaj, II = Ojdx j, then the LHS becomes X(O;) = XjajO" and the RHS can be written as

(LxII),

+ (II, [X j aj, ail). ~

-(ajxi)aj

It follows that

(26.33) We give two other useful properties of the Lie derivative applicable to all tensors. From the Jacobi identity one can readily deduce that

L[x,YlZ = LxLyZ - LyLxZ. Similarly, Equation (26.32) yields

L[X,y]1I = LxLyll- Ly LxII Putting these two equations together, recalling that a general tensor is a linear combination of tensor products of vectors and I-forms, and that the Lie derivative obeys the product rule of differentiation, we obtain

L[X,y]T = LXLyT - LyLxT

(26.34)

for any tensor field T. Furthermore, Equation (26.33) and the linearity of the Lie bracket imply that L.x+py = aLx + fJLy when acting on vectors and I-forms. It follows by the same argument as above that

L.X+pyT = aLxT + fJLyT

v

TE'1;(M).

(26.35)

26.5 EXTERIOR CALCULUS

26.5

791

Exterior Calculus

Skew-symmetric tensors are of special importance to applications. We studied these tensors in their algebraic format in the last chapter. Let us now investigate them as they reside on manifolds. differential forms. or simply forms, defined

26.5.1. Definition. Let M be a manifold and Q a point of M. Let A~(M) denote the space ofall antisymmetric tensors of rank p over the tangent space at Q. Let AP(M) be the union of all A~(M) for all Q E M. A dijJerential p-form w is a mapping w : U -> AP(M) such that w(Q) E A~(M) where U is, as usual, a subset of M. To emphasize their domain of definition, we sometimes use the notation A P(U). Since {dxi}7'~l is a basis for TQ(M) at every Q E M, {dx i1 A ... A dx ip} is a basis for the p-forms. All the algebraic properties established in the last chapter apply to these p-forms at every point Q E M. The concept of a pullback has been mentioned a number of times in connection with linear maps. The most frequent use of pullbacks lakes place in conjunction with the p-forms.

defining pullback for differential forms

26.5.2. Definition. Let M and N be manifolds and 1/r : M -> N a differentiable map. The pullback map on p-forms is the map 1/r* : AP(N) -> AP(M) defined by

for P For p = 0, i.e.i for functions on M, 1/r*w sa w

0

E

AP(N).

1/r.

It can be shown that

(1/r 0



1/r*.

(26.36)

Since w varies from point to point, we can define its derivatives. Recall that Tg(M) is the collection of real-valued functions on M. Since the dual of R is JR, we conclude that A o(M), the collection of zero-forms, is the union of all real-valued functions on M. Also recall that if f is a zero-form, then df , the differential of f, is a one-form. Thus, the differential operator d creates a one-form from a zero-form. The fact that this can be generalized to p- forms is the subject of the next theorem (for a proof, see [Abra 88, pp. 111-112]). exterior derivative and itsantiderivation property

26.5.3. Theorem. For each point Q of M, there exists a neighborhood U and a unique operator d : AP (U) -> AP+! (U), called the exterior derivative operator, such thatfor any WE AP(U) and rz E Aq(U),

1. dew

+ rz) =

dw + dn ifq = p; otherwise the sum is not defined.

2. dew A rz) = (dw) A rz + (-I)Pw A (drz); this is called the antiderivation property ofd with respect to the wedge product.

792

26. ANALYSIS OF TENSORS

3. d(dwl = afar any differential form w; stated differently, dod = O. 4. df = (a;jldx i for any real-valuedfunction f.

5. d is natural with respect to pullback; that is, du 0 1ft* = 1ft* 0 dN for any differentiable map 1ft : M ...,. N. Here du (dNi is the exterior derivative operating on differential forms of M (N i. 26.5.4. Example. Let M ofwis

= IR3 and w =

ajdx i a l-form on M. The exterior derivative

dW=(dai)Adxi = (ajaidxj)Adx i = ~)ajai -aiaj)dxj Adx i. j-ci

Wesee that thecomponents of dw arethecomponents of V x A where A = Cal, oz. a3). Itfollows that thecurlof a vector in JR.3 is theexterior derivative of the l-form constructed outof the components of thevector. 1m The homogeneous Maxwell's equations are written in terms of differential forms.

26.5.5. Example. In relativistic electromagnetic theory theelectric and magnetic fields are combined toformtheelectromagnetic fieldtensor. Thisis askew-symmetric tensor field of rank 2, whichcanbe written as8

F = -Exdt A dx - Eydt A dy - Ezdt A dz + Bzdx A dy - Bydx A dz + Bxdy Adz,

(26.37)

where t is the timecoordinate andtheunits aresuch that c, thevelocity of light, is equalto 1. Let us take the exterior derivative ofF. In theprocess, we use df = (a;f)dx i ,d(dx i A dx j) = 0, and in dE; or dBj we include only the tenus that give a nonzero contribution:

(aE y dx + aE y dz) Adt Ady b h z Z Z _ (aE dx + aE dy) A dt A dz + (aB dt + aBz dz) A dx A dy ax ay at az y x _ (aB dt + aBy dy) A dx A dz + (aB dt + aBx dX) A dy Adz. at ay at ax

dF = - (aEx dy ~

+ aEx dz) A dt A dx _ h

Collecting all similar terms and taking intoaccount changes of signdueto theantisymmetry of theexterior products gives

aE y dF = ( -b

ee; - -aBy) dt A dx A dz aEx + -aBZ) dt A dx A dy + (aEZ - -- - -~ h b h at

aEZ aEy aBX ) + (- + dt A dy ~

h

h

1\

(aBx aBy aBZ) dz + + + dx A dy Adz b ~ az

[(v x E+ ~~)Jdt rc dx Ady + [(v x E+ ~~)Jdt Adz rc dx + [(v x E+ ~~)Jdt Ady Adz+ (V· B)dx dy Adz =

/v

SNotehow in the wedge product, the first factor has a lower index (is an"earlier" coordinate) than the secondfactor. If this restriction is to be removed, we need to introduce a factor of! foreach component (see Example26.5.11).

26.5 EXTERIOR CALCULUS

793

Eachcomponent of dF vanishes because of Maxwell's equations.

III

The example above shows that

26.5.6. Box. The two homogeneous Maxwell's equations can be written as dF = 0, where F is defined by Equation (26.37).

The exterior derivative is a very useful concept in the theory of differential forms, as illustrated in the preceding example. However. that is not the only differentiation available to the differential forms. We have already defined the Lie derivative for arbitrary tensors. Since differential forms are (antisymmetrized) linear combinations of covariant tensors, Lie differentiation is defined for them as well. In fact, since differential forms have no contravariant parts, one uses the pullback map F,* in the definition of the Lie derivative instead of F,-;' 1.

* -1 -F, d *w. Lxw = (F,) dt

(26.38)

The two derivatives defined so far have the following convenient property, whose proof is left as an exercise for the reader: 26.5.7. Theorem. The exterior derivative d is natural with respect to Lx (or commutes with l-x) for X E X(M); that is, d 0 Lx = Lx: 0 d.

In our definition of exterior product in the previous chapter, we assumed that the number ofvectors was equal to the number of linear functionals taken from the dual space [see Equation (25.9)]. As a result of this complete pairing, we always ended up with a number. It is useful, however, to define an "incomplete" pairing in whicb the number of vectors and dual vectors are not the same. In particular, if we have a p-fonn and a single vector field, then we can pair the vector field with one of the factors of the p-fonn to get a (p - I)-fonn. This process is important enough to warrant the following: interior product of a vector field and a differential form

26.5.8. Definition. Let X be a vector field and w a p-form on a manifold M. Then define ix: AP(M) ~ AP-l(M) by iXW(Xl, ... , Xp-l) = w(X, Xl, ... , Xp-l).

If w E Ao(M), i.e., if w is just a function, we set ixw = O. ixw is called the interior product or contraction of X and w. Another notation commonly usedfor ixw isXJw. Although no signature of "differentiation" appears on ix, it does have such a property:

794

26. ANALYSIS OF TENSORS

26.5.9. Theorem. Let w be a p-form and 1/ a q-form on a manifold M. Then, ix is an antiderivation with respect to the wedge product: ix(w A 1/) = (ixw) A 1/ + (-I)P w A (iX'l).

Proof. The proof follows directly from Definitions 25.3.3 and 26.5.8, and is left for the reader.

D

We have introduced three types of derivation on the algebra of differential forms: the exterior derivative, the Lie derivative, and the interior product. The following theorem connects all three derivations in a most useful way (see A[Abra 88, pp. 115-116]): Relation between d, Lx, and ix

26.5.10. Theorem. Let w E AP(M), f E AO(M), and X E X(M). Let ix : AP(M) ---> AP-l(M), d : AP(M) ---> AP+l(M), and Lg : AP(M) ---> AP(M)

be the interiorproduct, the exteriorderivative, and the Lie derivative, respectively. Then 1. ixdf = Lxf. 2. Lx = ix 0 d

+d 0

ix.

3. Lfxw = fLxw+df Aixw. If X = Xi 8j and w = Wili2 ...ip+ldxil A dX i2 1\ ... A dx i p+1 , then the reader may verify that iXfd = XiWiil ...ipdxil A··· A dxip. In particular, we have the useful formula

ix(dx il A dX i2 A ... A dxip+I) = Xi aj.li.z...i~+ldxh A. dx h A ... A dx i p

(26.39)

} }l ..• l p

= analysis ofthe Lorentz force law In the language of forms

x! (", ,il ,i2 ,ip+I ) d X h ~ €1t u (j ) u (h ) ... utr(jp) 1r

1f

A

d x jz A··· /\ d X j p.

26.5.11. Example. Let p = Padxa be the momentum one-form and write the electromagnetic field tensor as9 F = !Fa/3dx U A dx fJ , where a and f3 run over the values 0, 1,2, and 3 with 0 being the time index. Let dp dt

== (dPa)

dx a

dr

be the derivative of momentum with respect to the proper time, r. Also, let u = uP8p be the velocity four-vector of a charged particle. Then the Lorentz force law can be written simpty as dp/dr' = qF(u) == -qiuF, wbere q is the etectric charge of the particle whose 4-velocityis u. Note that F. a two-form, contracts with u, a vector, to give a one-form on 9The factor ~ is introduced here to avoid restricting the sum over a and

p.

26.5 EXTERIOR CALCULUS

795

the RHS. Thus, both sides are of the same type. Let us writethis equation in component form:

d;'u dxu

= -q!Fupiu(dxU AdxP) = -tqFup(uY8~~dx") =

tqFupuY(8~8t - 8~8~)dx"

= tqFup(uPdxU - uUdx P)

(26.40)

= tq(Fup - Fpu)uPdx U = (qFupuP)dx u. Equating thecompnnents on both sides,we get dpu/dT = qFupu P, wbicbmaybefamiliar to thereader. Tomake theequation evenmore familiar, consider thecomponent a = 1, dPl • 0 2 3 aT = qFlpU" = q[FlOu + F12U + F13 U i.

(26.41)

and recallthat uCl = dx Ol[d-e, where (dT)2 = (dt)2 _ (dx l)2 _ (dx 2)2 _ (dx 3)2 = (dt)2(1_ v 2)

and v = (dx1/dt, dx 2/dt, dx 3/dt) is the 3-velocity of the particle.Since x O = t, we get

°

I

u =dt dT

u

i

i

dx dT

Vi

fori

=-=,,~=

)1- v2

= 1,2,3.

Substituting this in (26.41) and remembering that FlO = -FQ1 = El' Fl2 = B3' and F13 = - F3l = - B2' we nbtain dpl dt)l- v 2

-~~'"

=q

[I El

)1- v 2

+ B3

V2

)1 - v 2

- B2

vs )1 - v2

]

,

or dp ; at = q[El + (V2 B3 -

V3 B2)] = [q(E + v x B)lt·

Theother components areobtained similarly. Thus, in vector form we have dp dt

.

= q(E + v x B),

where p now represents the 3-rnomentum of the particle. Thisis the Lorentz forcelaw for electromagnetism in its familiarform. Again,notethesimplification offered bythelanguage of fnrms. III

A combination that is very useful is that of the exterior derivative and the Hodge star operator. Recall that the latter is defined by

* (dx". A··· A dx '.p ) =

1 I} l ' . E.... p. dx'r" A···/\ dx'»; (m _ p)! lp+l ... lm

where m is the dimension of the martifold.

(26.42)

796

26. ANALYSIS OF TENSORS

26.5.12. Example. Let us calculate *F and d(*F) where F = !Fapdx a A dx P is the electromagnetic fieldtensor. Wehave *F = *(!Fapdxa Adx P) = !FaP * (dx a /\ dx P) =

~Fap~E~~dx" Adx"

and d(*F)

= d(~FaPE~~dx" Adx") = ~E~~Fap.ydxY

/\ dx"

A

dx",

where Fa[j,y ea aFap/ax Y. Wecan nowuse the components FjO = E], F12 = B3. F13 = -H2, and F23 = Ht to write d(*F) in terms of E and B. After a loug but straightforward calculation, we obtain d(*F) =

[(~~ - Vx B)J dt rcdx Ady + [(~~ - Vx B)J dt Adz rvdx + [(~~ -

V x

B)J dt Ady Adz+ (V .E)dx Ady Adz. (26.43)

The inhomogeneous pairof Maxwell'sequations is

aE

VxB=-+4rrJ,

at

V· E = 4rrp,

(26.44)

where p and J are charge andcurrent densities, respectively. Wecanputthesetwodensities = Jadx(X. Thus, together to form a four-current one-form with p as the zeroth component: J

Maxwell's inhomogeneous equations Inthe language offorms

* J = Ja(*dx(X) =

1 Jo 3!€~vpdX~

A

dx"

A

dx P

=~bA~A~+h&A~A~+~&A~Ab+~&AbA~

(26.45) = pdx Ady Adz - Pdt Ady Adz - JYdt Adz rvdx - JZdt r- dx Ady,

where we have used the facts that p = JO = Jo and J = (J x , JY, JZ) -(1x, J y, Jz).

III

Comparing Equations (26.43), (26.44), and (26.45), we note that

26.5.13. Box. In the language of forms, the inhomogeneous pair of Maxwell's equations has the simple appearance d(*F) = 4".(*.1).

Problem 26.15 shows that the relation d 2w = 0 is equivalent-at least in rn:.3_ to the vanishing of the carl of the gradient and the divergence of the cnrl. It is customary in physics to try to go backwards as well, that is, given that V x E = 0, to assume that E = V f for some function f. Similarly, we want to believe that V· B = 0 implies thatB = V x A.

26.5 EXTERIOR CALCULUS

closed and exact forms

regions thatare contractable to a point

converse ofthe Poincare lemma

797

What is the analogue of the above statement for a geueral p-form? A form w that satisfies dw = 0 is called a closed form. An exact form is one that can be written as the exterior derivative of another form. Thus, every exact form is automatically closed. This is the Poincare lemma. The converse of this lemma is true only if the region ofdefinition ofthe form is topologically simple, as explained in the following. Consider a p-form w defined on a region U of a manifold M. If all closed curves in U can be shrunk to a point in U, we say that U is contractable to a point. If w is not defined for a point P on M, then any U that contains P is not contractable to a point. We can now state the converse of the Poincare lemma (for a proof, see [Bish 80, p. 175]): 26.5.14. Theorem. (converse of the poincare lemma) Let U be a region in a manifold M, such that U is contractable to a point. Let w be a p-form on U such that dw = O. Then there exists a (p - I)-form '1 on U such that w = dn. 26.5.15. Example. The electromagnetic field tensor F = !Frtpdx ct /\.dx P is a two-form that satisfies dF = O. The converse of the Poincare lemma says that if F is well behaved in a region U of lR.4 , then there must exist a one-form TJ such that F = d'l]. Let us write this one-form in terms of coordinates as T1 = Aadx rt • Then dTJ = Aa,pdx fJ /\ dx", and we have

!F.pdx· f\dx P = Ap,.dx· f\ dx P

=}

!(F.p - Ap,.

+ A.,p)dx· f\ dx P = O.

Since dx" A dx fJ are linearly independent and their coefficients are antisymmetric, each of the latter must vanish. Thus,

The four-vector Aa is simply the four-potential ofrelativistic electromagnetic theory.

gauge invariance in the language of forms

II

Note that the (p - I)-form of Theorem 26.5.14 is not unique. In fact, if a is any (p - 2)-form, then w can be written as w = d('1 + da) because d(da) is identical to zero. This freedom of choice in selecting '1 is called gange invariance, and its generalization plays an important role in the physics of fundamental interactions.

Jules Henri Poincare (1854-1912): The development of mathematics in the nineteenth century began under the shadow of a giant, Carl FriedrichGauss; it ended with the domination by a genius of similar magnitude, Henri poincare. Both were universal mathematicians in the supreme sense, and both made important contributions to astronomy and mathematical physics. If Poincare's discoveries in number theory do not equal those of Gauss, his achievements in the theory of functions are at least on the same level-even when one takes

798

26. ANALYSIS OF TENSORS

into account the theory of elliptic and modular functions, which must he credited to Gauss and which represents in that field his most important discovery, although it was not published during his lifetime. If Gauss was the initiator in the theory ofdifferentiable manifolds, poincare played the same role in algebraic topology. Finally, Poincare remains the most important figure in the theory of differential equations and the mathematician who after Newton did the most remarkable work in celestial mechanics. Both Gauss and poincare had

very few studentsand liked to work alone;but the similarityends there. Where Gauss was very reluctant to publish his discoveries, Poincare's list of papers approaches five hundred, which does not include the many books and lecture notes he published as a result of his

teachingat the Sorbonne. Poincare's parents both belonged to the upper middle class, and both their families had lived in Lorraine for several generations. His paternal grandfather had two sons: Leon, Henri's father, was a physician and a professor of medicine at the University ofNancy; Antoine had studied at the Ecole Polytechnique and rose to high rank in the engineering corps. One ofAntoine's sons, Raymond, was several times prime minister and was president of the French Republic dnring World War I; the other son, Lucien, occupied high administrative functions in the university. Poincare's mathematical ability became apparent while he was still a student in the lycee, He won first prizes in the concours genial (a competition among students from all French lycees) and in 1873 entered the Ecole Polytechnique at the top of his class; his professor at Nancy is said to have referred to him as a "monster of mathematics." After graduation, he followed courses in engineering at the Ecole des Mines and worked briefly as an engineer while writing his thesis for the doctorate in mathematics which he obtained in 1879. Shortlyafterward he started teaching at the University of Caen, and in 1881 he became a professor at the University of Paris, where he taught until his untimely death in 1912. At the early age of thirty-three he was elected to the Academte des Sciences and in 1908 to the Academic Francaise. He was also the recipient of innumerable prizes and honors both in France and abroad. Before he was thirty years of age, poincare became world famous with his epoch-making discovery of the "automorphic functions" of one complex variable (or. as he called them, the "fuchsian" and "kleinean" functions). Much has been written on the "competition" between C. F. Klein and poincare in the discovery of automorphic functions. However, Poincare's ignorance of the mathematical literature when he started his researches is almost unbelievable. He hardly knew anything on the subject beyond Hermite's work on the modular functions; he certainly had never read Riemann, and by his own account had not even heard of the "Dirichlet principle," which he was to use in such imaginative fashion a few years later. Nevertheless, Poincare's idea of associating a fundamental domain to any fuchsian group does not seem to have occurred to Klein. nor did the idea of "using" non-Euclidean geometry, which is never mentioned in his papers on modular functions up to 1880. Poincare was one of the few mathematicians of his time who understood and admired the work of Lie and his continuators on "continuous groups." and in particular the only mathematician who in the early 1900s realized the depth and scope of E. Cartan's papers. In 1899 Poincare proved what is now called the Poincare....Birkhoff-Witt theorem which has become fundamental in the modem theory of Lie algebras. The theory of differential

26.5 EXTERIOR CALCULUS

799

equations and its applications to dynamics was clearly at the center ofPoincare's mathematical thought; from his first (1878) to his last (1912) paper, he attacked the theory from all possible angles and very seldom let a year pass without publishing a paper on the subject. The most extraordinary production of Poincare's, also dating from his prodigious period of creativity (1880-1883) (reminding us of Gauss's Tagebuch of 1797-1801), is the qualitative theory of differential equations. It is one of the few examples of a mathematical theory that

sprang apparently out of nowhere and that almost immediately reached perfection in the hands of its creator. Everything was new in the first two of the four big papers that poincare published on the subject between 1880 and 1886. For more than twenty years poincare lectured at the Sorbonne on mathematical physics; he gave himself to that task with his characteristic thoroughness and energy, with the result that he became an expert in practically all parts of theoretical physics, and published more than seventy papers and books on the most varied subjects, with a predilection for the theories of light and of electromagnetic waves. On two occasions he played an important part in the development of the new ideas and discoveries that revolutionized physics at the end of the nineteenth century. His remark on the possible connection between X-rays and the phenomenon of phosphorescence was the starting point of H. Becquerel's experiments that led him to the discovery of radioactivity. On the other hand, poincare was active from 1899 on in the discussions concerning Lorentz's theory of the electron; Poincare was the first to observe that the Lorentz transformations form a group; and many physicists consider that poincare shares with Lorentz and Einstein the credit for the invention of the special theory of relativity. The main leitmotiv of poincare's mathematical work is clearly the idea of "continuity": Whenever he attacks a problem in analysis, we almost immediately see him investigating what happens when the conditions of the problem are allowed to vary continuously. He was therefore bound to encounter at every tum what we now call topological problems. He himself said in 1901, "Every problem I had attacked led me to Analysis situs," particularly the researches on differential equations and on the periods of multiple integrals. Starting in 1894 he inaugurated in a remarkable series of six paperswritten during a period of ten years-the modem methods of algebraic topology. Whereas poincare has been accused of being too conservative in physics, he certainly was very open-minded regarding new mathematical ideas. The quotations inhis papers show that he read extensively, if not systematically, and was aware of all the latest developments in practically every branch of mathematics. He was probably the first mathematician to use Cantor's theory of sets in analysis. Up to a certain point, he also looked with favor on the axiomatic trend in mathematics, as it was developing toward the end of the nineteenth century, and he praised Hilbert's Grundlagen der Geometric. However, he obviously had a blind spot regarding the formalization of mathematics, and poked fun repeatedly at the efforts of the disciples of Peano and Russell in that direction; but, somewhat paradoxically, his criticism of the early attempts of Hilbert was probably the starting point of some of the most fruitful of the later developments of metamathematics. poincare stressed that Hilbert's point of view of defining objects by a system of axioms was admissible only if one could prove a priori that such a system did not imply contradiction, and it is well known that the proof of noncontradiction was the main goal of the theory that Hilbert founded after 1920. poincare seems to have been convinced that such attempts were hopeless, and K. Godel's theorem proved him right.

• 800

26. ANALYSIS OF TENSORS

26.5.1 Integration on Manifolds integration of ditierential forms in

JR"

We mentioned in Chapter 25 that certain exterior products are interpreted as volume elements. We now exploit this notion and define integration on manifolds. Starting with JR" , considered as a manifold, we define the integral of an n-form w as follows. Choose a coordinate system {x'}i~l in R", writew = fdx 1A· .. rcdx", and define ~ the integral of the n-form as

1.

co se

R",X

1.

f(x 1 , ••. , x")dx 1 ••. dx";



]RII

where to avoid dealing with infinities, one assumes that f vanishes outside a finite region. The second symbol in the lower part of the integral sigo indicates the variables of integration. Let us now change the coordinates, say to {yJ }j~l' Using Equation (26.17), which gives the transformation rule for I-forms when changing coordinates, and Equation (25.13), which defines the determinant in terms of nforms, we obtain r

(ax')

" w=fdx 1 A"'Adx" =fdet - . dy 1 A···Ady,

ayJ

,

where f is now understood to be a function of the y's through the x's. So, in terms of the new coordinates, the integral becomes

1.

w=

RII,y

1 RII

ax' )

" f(x 1(y), ... ,xn (y))det ( - . dy 1 A···Ady. 8y)

If we had the absolute value of the Jacobian in the integral, the two sides would be equal. So, all we can say at this point is

1.

R",y

This discussion is analogous to our discussion of

orientation in vector spaces (seeSection 25.3.2).

orientable manifolds

w=

±1.

Rn,x

w.

We therefore distinguish between two kinds of coordinate transformations: If the Jacobian determinant is positive, we say that the coordinate transformation is orientation preserving. Otherwise, the transformation is called orientation reversing. Our ability to integrate functions on JR" depends crucially on the fact that volume elements do not change sign at any point of JR". If this were not so, we could find a finite (albeit small) region of space-in the vicinity of the point at which the volume element changes sign-whose volume would be zero. This property of JR" is the content of the following: 26.5.16. Definition. A manifold M ofdimension n is called orientable ifit has a nowhere vanishing n-form.

26.6 SYMPLECTIC GEOMETRY

801

Any two nonvanishing n-forms w and w' on an orientable manifold are related by a nowhere-vanishing function: w' = fw. Clearly, f has to be either positive or negative everywhere. w and w' are said to be equivalent if f is positive. Thus, the nonvanishing n-forms on an orientable manifold fall into two classes, all members of each class being eqnivalent to one another, and a member of one class being related to a member of the other class via a negative function. Each class is called an orientation on M. Given an orientation, an n-form w, and a chart {Va, M, so that it maps n-forms on M to n-fonns on R"; wl a is the restriction ofw to U«, andthe sum over a is assumedto

"volume" ofa manifold

exist. This amounts to saying that the region in M on which w is defined is finite, or that w has compact support. We note that the RHS of Equation (26.46) is an integration on JR" that appears to depend on the choice of coordinate functions. However, it can be shown that the integral is independent of such choice. In practice, one chooses a coordinate patch and transfers the integration to JR", where the process is familiar. lfwe choose a coordinate patch {xili'=1 and integrate dx 1/\ ... r.dx" according to Equation (26.46), we obtain the "volume" of the manifold M. lf M is compact, this volume will be finite. 10

26.6 Symplectic Geometry Mechanics stimnlated a great deal of dialogue between physics and mathematics in the latter part of the nineteenth century and the beginning of the twentieth. The branch of mathematics that benefited the most out of this dialog is the theory of differentiable manifolds, whose tribute back to mechanics has been the most beautiful language in which the latter can express itself, the language of symplectic geometry. All the discussion of symplectic vector spaces of the last chapter can be carried over to the tangent spaces of a manifold and patched together by the differentiable sttucttue of the manifold. symplectic form, symplectic structure, and symplectic manifold defined

26.6.1. Definition. A symplectic form (or a symplectic structure) on a manifold M is a nondegenerate, closed 2-form w on M. A symplectic manifold (M, w) is a manifold M together with a symplectic form w on M. We define the map b : X(M) ---> X*(M) by b(X) ea X'

= ixw = w' (X)

lORecall fromChapter 16 thata subsetof:Rn is compact iff it is closed andbounded. It is a good idea to keep this in mindas a paradigm of compactspaces.

802

26. ANALYSIS OF TENSORS

and the map

~

: X*(M) --> X(M) as the inverse of».

Chapter 25 identified some special basis, the canonical basis, in which the symplectic form of a symplectic vector space took on a simple expression. The analogue of such a basis exists in a symplectic manifold. The reader should keep in miud that this existence is not automatic, because although one can fiod such bases at every point of the manifold, the smooth patching up of all such bases to cover the entire manifold is not trivial and is the content of the following important theorem, which we state without proo~ (see [Abra 85, p. 175]): Darboux theorem

26.6.2. Theorem. (Darboux) Suppose w is a 2-form on a 2n-dimensional manifold M. Then dw = 0 if and only if there is a chart (U, rp) at each P E M such that rp(P) = 0 and

" i w = Ldx AdJ', ;=1

where Xl , .•• , x", yl, ... , yn are coordinates on U.Furthermore, on such a chart, the volume element ttw is fL", = dx 1 A ... A dx" A dyl A ... A dy",

symplectic charts, canonical coordinates, and canonical transformations Coordinate representation of sharpandflat maps

26.6.3. Definition. The charts guaranteed by Darboux's theorem are called symplectic charts, and the coordinates xi, yi are called canonical coordinates. If (M, w) and (N, p) are symplectic manifolds, then a eoo map f : M --> N is called symplectic, or a canonical transformation, if f* P = w. 26.6.4. Example. In this example, we derive a formula that gives the action of wU in termsof components of vectorsand I-forms in canonical coordinates. Let

.a

Z=X'-.

ax'

w' and

.a

+Y'-. ay'

be a vector field. When w D acts on Z, it gives a 'l-form, which we write as wD(Z) _ Ukdxk + Wkdi· To find the nnknowns Uk and Wko we let both sides act on coordinate basis vectors.Forthe RHS, we get

and

26.6 SYMPLECTIC GEOMETRY

803

For the LHS, we obtain

But

and

It follows that

. a +yl_. . a )] ( _. a ) =_yJ.. [w ,( X'-. ox' oy' ax) Similarly,

Therefore, (26.47)

where a summation overrepeated indices is understood. If we multiply both sides of this equation by w Uou the left and recall that wUw' = 1,

we obtain the following equation for theaction of wU: (26.48) Equations (26.47) and (26.48) are very useful in Hamiltonian mechanics.

III

804

26. ANALYSIS OF TENSORS

Our discussion of symplectic transformations of symplectic vector spaces showed that such maps are necessarily isomorphisms. Applied to the present situation, this means that if f : M -> N is symplectic, then f* : 'Jp (M) -> 'J f(P) (N) is an isomorphism. Theorem 26.3.2, the inverse mapping theorem, now gives the following theorem. 26.6.5. Theorem. If f : M -> N is symplectic, then it is a local diffeomorphism.

from Lagrangian to Hamiltonian Inthe language of differenfial forms

26.6.6. Example. Hamiltonian mechanics takes place in the phase space of a system. Thephasespaceis derived fromtheconfiguration spaceas follows. Let (ql, ... , qn) be the generalized coordinates of a mechanical system. They describe ann-dimensional manifold N. The dynamics of the systemis describedby the (time-independent) LagrangianL, which is a functionof (qi, qi), Butqi are the componentsof a vectorat(ql."" qn) [seeEquation (26.13)and replace r' with xiI. Thus, in the language of manifoldtheory, a Lagrangianis a functionon the tangentboodle, L : T(N) -+ JR. The Hamiltonian is obtained fromtheLagrangian by a Legendre transfonnation: H = L:f=l Piiji - L. The first term can be thought of as a pairing of an element of the tangent space with its dual. In fact, if P has coordinates (q" ... , qn), then q == ai E 'Jp(N) (withtheEinstein summation convention enforced), andif we pairthiswiththedualvector Pjdx j E 'Jj.(N), we obtain the first term in the definitionof the Hamiltonian. The effect of the Legendre transformation is to replace by Pi as the second set of independent vatiables.This has the effectof replacing T(N) with T*(N). Thns

t/

il

26.6.7. Box. The manifold cfHamittontan dynamics, or the phase space, is T*(N), with coordinates (qi, Pi) on which the Hamiltonian H : T*(N) -)- R is defined.

symplectic 2-form of P(Mj

T*(N) is 2n-dimensional; so it has the potential of becominga symplectic manifold. In fact, it can be shownthatl l the 2-fonn suggestedby Darboux's theorem, n

w= Ldqi Adpi,

(26.49)

i=1

is nondegenerate, andtherefore a symplectic form for T*(N).

iii

The phase space, equipped with a symplectic form, torns into a geometric arena in which Hamiltonian mechanics unfolds. We saw in the above example that a Hamiltortian is a function on the phase space. More generally, if (M, w) is a symplectic manifold, a Hamiltortian H is a real-valned function, H : M -> JR. Given a Hamiltortian, one can define a vector field as follows. Consider dH E T* (M). For a symplectic manifold, there is a natural isomorphism between T* (M) 11Here, we are assuming that the mechanical system in question is nonsingular, by which is meantthat there areprecisely independent Pi's. There aresystems of considerable importance that happen to be singular. Such systems, amongwhichare included all gaugetheories suchas the general theory of relativity, arecalledconstrained systems andarecharacterized by the factthat w is degenerate. Although of great interest andcurrently under intense study, we shallnot discussconstrained systems in thisbook.

n

805

26.6 SYMPLECTIC GEOMETRY

and T (M), namely,w~. Theuniquevectorfield XH associated withdH is thevector field we are after. Hamiltonian vector field and Hamiltonian systems defined

26.6.8. Definition. Let (M, w) be a symplectic manifold and H : M --> R a real-valuedfunction. The vector field XH sa w~(dH) es (dH)~ is called the Hamiltonian vector field with energy function H. The triplet (M, w, XH) is called a Hamiltonian system.

The significance of the Hamiltonianvector field lies io its iotegraicurve which turns out to be the path of evolutionof the system io the phase space.This is shown io the followiog proposition. 26.6.9. Proposition. If (ql, ... , e". PI, ... , Pn) are canonical coordinates for w-so w = L dq' /\ dpi-then, in these coordinates

a~i == (:;, - :;) .

(26.50)

Therefore, (q(t), p(t)) is an integral curve ofXH

iff Hamilton's equations hold:

XH = : ; a: i - : ;

aqi

at =

aH aP,'

api

at =

en - aqi '

i = 1, .. . ,n.

(26.51)

Proof. The first part of the proposition follows from

en

dH = -.dq aq'

i

aH + -dpi. api

from the definitionof XH io terms of dH, and from Equation (26.48). The second D part follows from the definition of iotegrai curve and Equation (26.19). We called H the energy function; this is for good reason: 26.6.10. Theorem. Let(M, w, XH) be a Hamiltonian system andy (r) an integral energy in the curve ofXH. Then H(y(t)) is constant in t. language of symplectic geomelry Proof. We show that the time-derivative of H (y (t)) is zero: conservation of

d

dt H(y(t)) = Y.t(H)

by Proposition 26.2.4

= dH(Y.t) = dH (XH(y(t)))

by Equation (26.14) by definitionof iotegrai curve

= [w' (XH(y(t)))] (XH(y(t)))

by definitionof XH(y(t))

= w (XH(y(t)), XH(l'(t)))

by the definitionof because w is skew-symmetric

=0

w'

D

806

26. ANALYSIS OF TENSORS

Theorem 26.6.10 is the statement of the conservation of energy. In the theoretical development of mechanics, canonical transformations play a central role. The following proposition shows that the flows of a Hamiltonian

system are such transformations: flow of Hamiltonian vectorfield is canonical transformation of mechanics

26.6.11. Proposition. Let (M, w, XH) be a Hamiltonian system, and F, the flow OfXH. Thenfor each t, F,*w = w, i.e., F, is symplectic.

Proof We have by Equation (26.38)

= Ft' (ixHdw + dixH w) = F,*(O

=0

+ ddH)

by Theorem 26.5.10 because dw because d 2

= 0 and ixw = w b(X) ~0

Thus, Ft*w is constant in t. But Fa = id. Therefore, Ftw = w.

D

Sir William Rowan Hamilton (1805-1865), the fourth of nine children, was mostly raised by an uncle, who quickly realized theextraordinary nature of his youngnephew. By the age of five, Hamilton spoke Latin, Greek, and Hebrew, and by the age of nine had added morethana half dozen languages to that list. He was also quite famous for his skill at rapid calculation. Hamilton's introduction to mathematics carne at the age of 13, when he studied Clairant's Algebra, a task made somewhat easier as Hamilton was fluent in French by this time. At age 15 he started studying Newton, whose Principia spawnedan interest in astronomy that would provide a great influence in Hamilton's earlycareer. In 1822, at the age of 18, Hamilton entered Trinity College, Dublin, and in his first yearhe obtained the top mark in classics. He dividedhis studies equallybetweenclassics andmathematics andin his secondyearhe receivedthetop award in mathematical physics. Hamilton discovered an error in Laplace's Mechanique celeste, andas a result, he cameto the attention of JohnBrinkley, the Astronomer Royal of Ireland, who said: ''This young man, I do not say will be, butis, the first mathematician of his age." While in his finalyear as anundergraduate, he presented a memoir entitledTheory ofSystems ofRays to theRoyal Irish Academy in whichhe planted the seeds of symplectic geometry. Hamilton's personal life was marked at first by despondency. Rejected by a college friend's sister, he became ill andnearlysuicidal. He wasrejected afew years later by another friend's sisterandwoundup marrying a very timidwomanprone to ill health. Hamilton's own personality was much more energetic andhumorous, andhe easily acquired friends among the literati. His own attempts at poetry, which he himself fancied, were generally considered quite poor. No less an authority than Wordsworth attempted to convince him

26.6 SYMPLECTIC GEOMETRY

807

that his true calling was mathematics, not poetry. Nevertheless, Hamilton maintained close connection with the worlds of literature and philosophy,insisting that the ideas to be gleaned from them were integral parts afhis life's work. While Hamilton is best known in physics for his work in dynamics, more of his time was spent on studies in optics and the theory of quaternions. In optics, he derived a function of the initial and final coordinates of a ray and termed it the "characteristic function," claiming that it contained "the whole ofmathematical

optics."Interestingly,his approach shed no new light on the wave/corpusculardebate (being independentof which view was taken), another appearance of Hamilton's quest for ultimate generality. In 1833 Hamilton published a study of vectors as ordered pairs. He used algebra to study dynamics in On a General Method in Dynamics in 1834. The theory of quaternions, on which he spent most of his time, grew from his dissatisfaction with the current state of the theoretical foundation of algebra. He was aware of the description of complex numbers as points in a plane and wondered if any other geometrical representation was possible or if there existed some hypercomplex number that could be represented by three-dimensional points in space. If the latter supposition were true, it would entail a natural algebraic representation of ordinary space. To his surprise, Hamilton found that in order to create a

hypercomplexnumber algebra for which the modnlus of a pradnct eqnaled the prodnct of the two moduli,four components were required-hence, quaremions. Hamilton felt that this discovery would revolutionize mathematical physics, and he spent the rest of his life working on quaternions, including publication of a book entitled Elements ofQuaternions, which he estimated would be 400 pages long and take two years to write. The title suggests that Hamilton modeled his work on Euclid's Elements and indeed this was the case. The hook ended up double its intended length and took seven years to write. In fact, the final chapter was incomplete when Hamilton died, and the book was finally published with a preface by his son, William Edwin Hamilton. While quaternions themselves turned out to be of no such monumental importance, their appearance as the first noncommutative algebra opened the door for much research in this field, including much of vector and matrix analysis. (As a side note, the "del" operator, named later by Gibbs, was introduced by Hamilton in his papers on quatemions.) In dynamics, Hamilton extended his characteristic function from optics to the classical action for a system moving between two points in configuration space. A simple trans-

formation of this functiou gives the quantity (the time integral of the Lagrangian) whose variation equals zero in what we now call Hamilton's principle. Jacobi later simplified the application of Hamilton's idea to mechanics, and it is the Hamilton-Jacobi equation that is most often used in such problems. Hamiltonian dynamics was rescued from what could have become historical obscurity with the advent of quantum mechanics, in which its close association with ideas in optics found fertile application in the wave mechanics of de Broglie and Schr6dinger. Hamilton's later life was unhappy, and he became addicted to alcohol. He died from a severe attack of gout shortly after receiving the news that he had been elected the first foreign member of the National Academy of Sciences of the USA.

Liouville's theorem

The celebrated Liouville's theorem of mechauics, concerniug the preservation of volume of the phase space, is a consequence of the proposition above: 26.6.12. Corollary. (Liouville's Theorem) F, preserves the phase volume I-L",.

808

26. ANALYSIS OF TENSORS

Poisson brackets in the language of symplectic geometry

26.6.13. Definition. Let (M, w) be a symplectic manifold. Let f, g : M -+ ffi. with X j = (df)U andXg = (dg)U their corresponding Hamiltonian vector fields. The Poisson bracket of f and g is the function

{f, g} sa w(Xj, Xg)

= ix,ixfw = -iXfix,w.

We can immediately obtain the familiar expression for the Poisson bracket of two functions.

26.6.14. Proposition. In canonical coordinates (qi, ... , q"; Pt, ... , Pn), we have

{f g} =

,

~ (a ~ _ .!L~) tSi aqi Bp, api aqi . f

In particular, {qi, qi} = 0,

{qi, Pi} = 8~.

{Pi, Pi} = 0,

Proof. From Equation (26.50), we bave

n

=I: i,j=l

::i aa:i

W

(a:i' a:i) - ::i ::i

'-----,,--'

W (a: i' a;i) '------.-'

=0

- ::i ::i

W

C~i'

a:i) + ::i aa:i

'-----,,--'

=-8~

=

=~f

(a~i'

W a;i)] '------.-'

=0

n (a f ag af ag) aqi api - api aqi '

~

where we have assumed that W = Lk=l dqk A dpi, The other formulas follow immediately once we substitute Pi or qi for f or g. D

26.7 Problems 26.1. Provide the details of the fact that.a finite-dimensional vector space V is a manifold of dimension dim V. ,

26.7 PROBLEMS

809

26.2. Chooseadifferentcurvey : R -+ W whose tangent at u = Oisstill(ux, Uy) of Example 26.2.2. For instance, you may choose y(u) = (~(u

+ 1)2, ~(u -

1)3).

Show that this curve gives the same relation between partials and unit vectors as obtained in that example. Can you find another curve doing the same job?

26.3. For every t

E 'J"p(M) and every constant function C E FOO(P), show that t(c) = O. Hint: Use both parts of Definition 26.2.3 on the two functions f = c andg=1.

26.4. Find the coordinate vector field at of Example 26.2.10. 26.5. Use the procedure of Example 26.2.10 to find a coordinate frame for S2 corresponding to the stereographic projection charts (See Example 26.1.12). 26.6. Let (xi) and (yj) be coordinate systems on a subset U of a manifold M. Let Xi and yi be the components of a vector field with respect to the two coordinate systems. Show that yi = xt ayi /ax j 26.7. Show that if Vr : M -+ N is a local diffeomorphism at P EM, then Vr.P : 'J"p(M) -+ 'J"",(p)(N) is a vector space isomorphism.

26.8. Let X be a vector field on M and Vr : M -+ N a differentiable map. Then for any function f on N, [Vr.X](f) is a function on N. Show that X(f

0

Vr) =

([Vr.X](f)}

0

Vr.

26.9. Verify that the vector field X = -yax

+ xay has an integral curve through

(xo, YO) given by

x = XQcost - yosint, Y = xosint + yocost. 26.10. Show that the vector field X = x 2ax

+ xya y has an integral curve through

(xo, YO) given by xo x(t)=--, I -xot

YO

y(t) = - - . I-xot

26.11. Let X and Y be vector fields. Show that X field, i.e., it satisfies the derivation property.

0

Y- X

0

Y is also a vector

26.12. Prove the remaining parts of Proposition 26.4.13. 26.13. Suppose that xi are coordinate functions on a subset of M and wand X are a l-form and a vector field there. Express w(X) in terms of component functions ofwandX.

810

26. ANALYSIS OF TENSORS

26.14. Show that d 0 Lx = Lx 0 d. Hint: Use the definition of the Lie derivative for p-forms and the fact that d commutes with the pullback. 26.15. Let M = ]R3 and let f be a real-valued function. Let w = a.dx' be a one-form and Tj = h 1dx 2 A dx 3 + h2dx3 A dx 1 + h3dx1 A dx 2 be a two-form on ]R3. Show that (a) df gives the gradient of f, (b) dn gives the divergence of the vector B = (hi, h2, h3), and that (c) V x (V f) = 0 and V . (V x A) = 0 are consequences of d 2 = O. 26.16. Show that ix: is an antiderivation with respect to the wedge product. 26.17. Given that F =

!Fa~dxa A dx~, show that F A (*F) = IB 2 1

IEI2 .

26.18. Use Equation (26.40) to show that the zeroth component of the relativistic Lorentz force law gives the rate of change of euergy due to the electric field, and that the maguetic field does uot change the energy. 26,19. Derive Equation (26.43).

!

26.20. Giveu that F = Fa~dxa A dx~, write the two homogeneous Maxwell's equations, V . B = 0 and V x E + aB/at = 0, in terms of Fa~. 26.21. Write the equatiou aA~ aA a Fa~ = A~,a - Aa.~ = axa - ax~

in terms of E, B, and vector and scalar potentials.

!

26.22. With F = Fa~dxa A dx~ and J takes the following form in components:

=

Jydx Y , show that d * F

= 4".(*J)

aFa~

- . - =4".r, ax" where indices are raised and lowered by diag(-1, -1, -1, 1). 26.23. Interpret Theorem 26.5.14 for p = 1 and p = 2 on ]R3. 26.24. Let f be a function on]R3. Calculate d * df. 26.25. Show that current conservation is an automatic consequence of Maxwell's inhomogeneous equation d * F = 4".(*J).

Additional Reading 1. Abraham, R. and Marsden, J. Foundations ofMechanics, 2nd ed., AddisonWesley, 1985. The definitive textbook on classical mechanics presented in the language of differentiable manifolds. Contains a thorough treatment of exterior calculus and symplectic geometry.

26.7 PROBLEMS

811

2. Bishop, R. and Goldberg, S. Tensor Analysis on Manifolds, Dover, 1980. 3. Bott, R. and Tu, L. Differential Forms in Algebraic Topology, SpriogerVerlag, 1982. Despite its frighteniog title, the first chapter of this book is actoally a very good iotroduction to tensors. 4. Choquet-Brnhat, Y., DeWitt-Morette, C., and Dillard-Bleick, M. Analysis, Manifolds, and Physics, 2nd ed., North-Holland, 1982. A two-volume reference written by mathematical physicists. Excellent for readers already familiar with the subject and seeking detailed applications io physics. 5. Warner, F. Foundations of Differentiable Manifolds and Lie Groups, Sprioger-Verlag, 1983. A formal but readable introduction to differentiable manifolds.

Part VIII

-----_

Lie Groups and Their Applications

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

27

_

Lie Groups and Lie Algebras

The theory ofdifferential equations had flourished to snch a level by the 1860s that a systematic study of their solutions became possible. Sophus Lie, a Norwegian mathematician, undertook such a study using the same tool that was developed by Galois and others to study algebraic equations: group theory. The groups associated with the study of differential equations, now called Lie groups, unlike their algebraic counterparts, are uncountably infinite, and, as such, are both intricate and full of far-reaching structures. It was beyond the wildest dream of any 19thcentury mathematician to imagine that a concept as abstract as Lie groups would someday find application in the study of the heart of matter. Yet, three of the four fundamental interactions are described by Lie groups, and the fourth one, gravity, is described in a language very akin to the other three.

27.1 Lie Groups and Their Algebras Lie groups are infinite groups that have the extra property that their multiplication law is differentiable. We have seen that the natural setting for differentiation is the structure of a manifold. Thus, Lie groups must have manifold properties as well as group properties. Lie groups defined

27.1.1. Definition. A Lie group G is a differentiable manifold endowed with a group structure such that the group operation G x G -+ G and the map G -+ G given by g >-+ g-l are differentiable. If the dimension ofthe underlying manifold is r, we say that G is an r-parameter Lie group.

Because of the dual nature of Lie groups, most of their mapping properties combine those of groups and manifolds. For instance, a Lie group homomor-

816

27. LIE GROUPS AND LIE ALGEBRAS

e

phism is a group homomorphism that is also oo , and a Lie group isomorphism is a group isomorphism that is also a diffeomorphism.

27.1.2. Example. GL(V) IS A LIE GROUP GL(V) is a Lie group

As the paradigm of Lie groups, we consider GL(V), the set of invertible operators on an n-dimensional real vector space '\7, and show that it is indeed a Lie group. The set L(V) is a vector space of dimension n 2 (Proposition 25.1.1), and Iherefore, by Example 26.1.7, a manifold of the same dimension. The map det : C(V) -+ IR is a oo map because the determinant, when expressed in terms of a matrix, is a polynomial. In particular, it is continuous. Now note that

e

GL(V) = deCI(1R - {O))

and that R - {O} is open. It follows Ihat GL(V) is an open submanifold of £,(V). Thus, G L(V) is an n 2-dimensional manifold.Choosinga basis B for Vand representingoperators (points) A of G L(V) as matrices (aij) in Ihat basis provides a coordinate patch for G L (V). We denote this coordinate patch by {xU}, where x ij (A) = aij' To show Ihat GL(V) is a Lie group, we need to prove Ihat if A, B E GL(V), Ihen AB : GL(V) x GL(V) -+ GL(V)

and

A-I: GL(V) -+ GL(V)

are e oo maps of manifolds. This is done by showing that the coordinate representations of . these maps are oo. These representations are simply the matrix representations of operators. Since AS is a linear function of elements of the two matrices, it has derivatives of all orders. It follows that AS is oo. The case of A-I is only slightly more complicated. We note that

e

e

A-I

= P(aU)

P(aij) = a polynomial in aij.

detA '

Thus, since detA is also a polynomial in aij, the kth derivative of A-I is of the form Q(aij)/(detA)k, where Q is anoIher polynomial. The fact that detA # 0 establishes Ihe oo property of A-I. One can similarly show that if V is a complex vector space, then GL(V) is a manifold of dimension 2n 2. III

e

27.1.3. Example. SL(V) IS A LIE GROUP SL(V) is a Lie group

Recall that SL(V) is Ihe subgroup of GL(V) whose elements have unit determinant. Since det: GL(V) -+ IR is oo, Theorem 26.3.7 and the example after it show Ihat SL(V) = det" 1(I) is a submanifold of GL(V) of dimension dim GL(V) -dim IR = n2 -I. Since it is already a subgroup, we conclude Ihat SL(V) is also a Lie group (Problem 27.5). Similarly, III when V is a complex vector space, one can show that dim SL(V) 2n 2 - 2

e

=

27.1.4. Example. OrnER EXAMPLES OF LIE GROUPS The reader may verify Ihe following: (a) Any finite-dimensional vector space is a Lie group under vector addition. (b) The unit circle SI, as a subset of nonzero multiplicative complex numbers is a Lie group under multiplication. (c) The product G x H of two Lie groups is itself a Lie group wiIh Ihe product manifold structure and the direct product group structure. (d) G L (n, JR.), the set ofinvertible n x n matrices, is aLie group under matrix multiplication.

27.1 LIE GROUPS ANO THEIR ALGEBRAS

group ofaffine motions of R n

817

(e) Let G = GL(n, R} x R" be the product manifold. Define the group operatiou by (A, u)(B, v) == (AB, Av + u). The reader may verify that this operatiou indeed defiues a group structure on G. In fact, G becomes a Lie group, called the group of affine motions of R", for if we identify (A,u) with the affine motion! x 1---* Ax + u of R", then the group operation in G is composition of affine motions. We shall study in some detail the poincare group, asubgroup of the group of affine motions, in which the matrices are (pseudo)

orthogoual.

local Lie groups

III

In calculations, one translates all group operations to the corresponding operations of charts. This is particularly useful when the group multiplication can be defined ouly locally. One then speaks of an r-parameter local Lie group. To be precise, one considers a neighborhood U of the origin of lll.' and defines an associative "multiplication" m : U x U ---+ R" and an inversion i : Uo ---+ U where Uo is a subset of U. We therefore write the multiplication as

m(a, b) = c,

a, b, C E IRY ,

where a = (0 1, 0 2, ... , or), etc. are coordinates of elements of G. The coordinates of the identity element of G are taken to be all zero. Thus, m(a, 0) = a and m(a, i(a)) = O. In component forms, k = 1,2, ... , r.

(27.1)

The fact that G is a manifold implies that all functions in Equation (27.1) are infinitely differentiable. 27.1.5. Example. As an example of a loca11-parameter Lie group, consider the multiE lll.llxl < I) and

plicationrule m : U x U -+ lll. where U = (x m(x, y) =

2xy -x - y I' xy -

x,y

E

U.

The reader can check that m(x, (y, z)) = m«x, y), z), so.that the multiplication is associative. Moreover,m(O,x) = m(x,O) = x for all x E U, andi(x) = x/(2x -i), defined forUO = (x E lll.llxl < III

!).

27.1.1 Group Action As mentioned in our discussion of finite groups, the action of a group on a set is more easily conceived than abstract groups. In the case of Lie groups, the natural action is not on an arbitrary set, but on a manifold.

local group of transformations

27.1.6. Definition, Let M be a manifold. A local group oftransformations acting on M is a (local) Lie group G, an (open) subset U with the property leJ x M C U c G x M, and a map \jI : U --+ M satisfying the following conditions:

IThese consist of a linear transformation followed by a translation.

818

27. LIE GROUPS AND LIE ALGEBRAS

M

For small regions of M. we may be able to include a large portion of G.

Figure 27.1

However, if we want to include all of M, as we should, then onlya smallneighborhood of the identitymaybe available.

1. If(g, P)

E

U, (h, 1jI(g, P))

E

U, and (hg, P)

E

U, then

1jI(h, W(g, P)) = W(hg, Pl. 2. Wee, P) = P for all P E M. 3. 1f(g, P) E U, then (g-I, 1jI(g, P)) E U and W(g-I, W(g, P)) = P. Normally, we shall denote W(g, P) by g. P, or g P, Then the conditions of the definition above take the simple fonn

g. (h . P) = (gh) . P, e P = P »

g-1 . (g . P) = P,

orbit; transitive action; stabilizer

for all

g, h

E

G,

P E M,

P EM, (27.2)

g E G, P EM,

whenever g . P is defined. Note that the word "local" refers to G and not M, i.e., we may have to choose a very small neighborhood of the identity before all the elements of that neighborhood can act on all points of M (see Figure 27.1). All the properties of a group action described in Chapter 23 can be applied here as well. So, one talks about the orbit of G as the collection of points in M obtained from one another by the action of G; a transitive action of G on M when there is only one orbit; the stabilizer of a point of M, etc. The only extra condition one has to be aware of is that the group action is not defined for all elements of G, and that a sufficiently small neighborhood of the identity needs to be chosen. In the old literature, the group action is described in terms of coordinates. Although for calculations this is desirable, it can be very clumsy for formal discussions, as we shall see later. Let a = (a l , ... , ar) be a coordinate system

819

27.1 LIE GROUPS AND THEIR ALGEBRAS

on G and x = (r ', ... , x") a coordinate system on M. Then the gronp action '11 : G x M --> M becomes a set of n fnnctions described by

x!

x"

= W(a, x),

= Web, x') = W(m(b, a), x),

(27.3)

where m is the mnltiplication law of the Lie gronp written in terms of coordinates as given in Eqnation (27.1). It is assnmed that'll is infinitely differentiable.

27.1.7. Box. Equation (27.3) can be used to unravel the multiplication law for the lie group when the latter is given in terms oftransformations.

27.1.8. Example. EXAMPLES OF GROUPS OF TRANSFORMATION (a) The two-dimensional rotation group acts on thexy-plane as <1>(0, r) = (x cos

If we write r'

y sinO, x sinO

+ y cosO).

= <1>(01, r) and r' = <1>(02, r'), then a simple calculation shows that

r" = (x cos(OI With r"

a-

+ 02) -

Y sin(O!

+ 02), x sin(O! + 02) + y cos(O! + 02)).

= (m(O!, 02); r), we recognize the "multiplication"1aw as m(OI, 02) = 01 + 02.

The orbits are circles centeredat the origin. (b) Let M = JRn , a a fixed vector in JR", and G Ij/(t,

translations

x) = x + ra,

=R

Define

Ij/ :

JR x JR" -+ JR" by

x e R", t ElR..

This group action is globally defined. The orbits are straight lines parallel to a. The group is the set of translations in thedirection a in IRn . Thereader mayverify that the "multiplication" law is addition of t's.

(c) Let G = lR+ be the multiplicative group of nonzeropositive real numbers. Fix real numbers aI, a'2•... , an, not all zero. Define the action of G on lRn by

The orbits are obtained by choosing a point in lRn andapplying G to it for different A'S. Theresultis a Curve in R". Forexample, if n = 2, ct'l = 1, andct'2 = 2, we get, as theorbit containing xo the curve YO 2

Y=2"x, X

o

which is a parabola going through the origin and the point (xO, yO). Note that the orbit

scale transformations one-dimensional projective group

containing the originhasonly one point. Thisgroup is called the group of scale transformations. Themultiplication law is ordinary multiplication of (positive) realnumbers. (d) Let G = JR4 acton M = JR by

820

27. LIE GROUPS AND LIE ALGEBRAS

The reader may verify that this is indeed the action of a group (catch where the condition i= Oisused!), andifx' =
ala4 '-aZa3

n (al bl x = (a3bl

+ a2b3)x + al b:2 + a2 b4 , + a4b3)x + a3b2 + a4 b4

so that the multiplication rule is

III

This group is called the one-dimensional projective group.

27.1.2 Lie Algebra of a Lie Group The group property of a Lie group G provides a natural diffeomorphism on G that determines a substantial part of its structure. lefltranslation, left-invariant vector fields, lefl-invariant forms, and their "right" counterparts

27.1.9. Definition. Let G be a Lie group and g E G. The left translation by g is a diffeomorphism L g : G -> G defined by Lg(h) = gh

V hE G.

A vector field eon G is called left-invariant iffor each g E G, itself; i.e} or

Lg.(e(h)) = e(gh)

V g, h

eis L.-related to « G.

The set ofleft-invariant vector fields on G is denoted by g. A l-form whose pairing with a left-invariant vector field gives a constant function on G is called a leftinvariant l-form. The right translation by g, Rg : G -> G, and right-invariant vector fields and l-forms are defined similarly. The reader may easily check that right and left translations commute: V g,h E G

(27.4)

Moreover, if wi, is a I-form on ~,(G), then w E A l(G), given by wig sa L;.IW I" is a left-invariant I-form:

independent of g. It is convenient to have a coordinate representation of L g •. The coordinate representation of L g is simply the multiplication law L g (h) = m (g, h), 2When there is no danger of confusion, we shall use ~(h) for ~Ih'

821

27.1 LIE GROUPS AND THEIR ALGEBRAS

where we have used the same symbol for coordinates as for group elements. Equation (26.11) can now be used to write the coordinate representation of L g*:

am !lah! am 2/ah!

am!2/ah' lah'] am

:

.

am'lah!

am'lah'

L g* --+

,

(27.5)

( where all the derivatives in the matrix are evaluated at (g, h). We have already mentioned in Chapter 26 that X( G) is an infinite-dimensional Lie algebra under the Lie bracket ''multiplication.'' In general, no finitedimensional subspace can be found that-in and of itself-is also an algebra.' However, Lie groups are an exception: 27.1.10. Proposition. Let G be a Lie group and g the set of its left-invariant vector fields. Then g is a real vector space, and the map : g ---> 'Je(G), defined by(~,) = e(e), isa linear isomorphism. Therefore, dim g = dim 'Je(G) = dim G. Furthermore, g is closed under Lie brackets; i.e., g is a Lie algebra.

Proof Ilis clear that g is areal vector space. If (e) = (1/) fore, 1/ E g, then

v g E G => e=1/. This shows that is injective. To show that e is surjective, suppose that v E 'Je(G) and define the vector field e on G by e(g) = Lg*(v) for all g E G. Then (e) = v and E g, because

e

L g*

0

e(h)

= L g* 0

Lh*(V)

= Lgh*(V) = e(gh) '" e(Lgh) = eo Lg(h).

This proves the first part of the proposition, The second part follows immediately from the definition of a left-invariant vector field and Theorem 26.4.4. D The flow of

eat g

E

G can be shown to be (27.6)

Ft = g exp(te) = Rexp(telg.

Indeed, let X, be the vector field associated with this flow. The action ofthis vector field on a function f is

X,l g (f) = :t

(f(gexPte))lt~o'

Therefore,

(Lh*X,lg)(f)

= X,lg (f 0

Lh)

=

d -d (f·0 Lh(gexpteJ) I t t=O

= dd (f(hgexpteJ)1

t

3Such a subspace is calleda subalgebra.

t=O

= X,(hg)(f) = [X; 0 Lh(g)](f).

822

27. LIE GROUPS AND LIE ALGEBRAS

Since this is true for all f and g, and X~ Ie = e(e), we conclude that X~ is the unique left-invariant vector field corresponding to e(e). The Lie algebra 01 a Lie group

27.1.11. Definition. The Lie algebra of the Lie group G is the Lie algebra 9 of left-invariant vector fields on G. Sometimes we think of as a vector in Te(G). In that case, we denote by X~ the left-invariant vector field whose value at the identity is e-

e

The isomorphism of 9 with Te(G) induces a Lie bracket on Te(G) and turns it into a Lie algebra. In many cases of physical interest, it is this interpretation of the Lie algebra of G that is most nseful. If two gronps stand in some algebraic relation to one another, their Liealgebras will inherit such relations. More precise! y, let G and H be Lie groups with Lie algebras 9 and ~, respectively. Suppose


v e, rz E g.

(27.7)

In particular, if c/> is a Lie group isomorphism, then c/>. is a Lie algebra isomorphism. 27.1.12. Example. Let V be a complexvectorspacewithits generallineargroup GL(V), a 2n 2-dimensional Lie group. Recall that GL(V) is an open submanifold of ,c,(V). By Equation (26.4), 'Je(GL(V)) = 'Je(,c,(V)), where e is the unit operator. If we identify 'Je(,c,(V)) with .c(V) [see the hox after Equation (26.4)], we may conclude that ,c,(V), which we deooteby g[(V) in the presentcontext,is the Lie algebraof GL(V) in whichthe Lie bracket is thecommutator. Weuse thenotation ACt) for a curve in GL(V) andAforthe vectortangentto the curve.

III

It is instructive to construct the coordinate representation of vector fields on GL(V). Let f : GL(V) --+ R be a function and A a vector field. Then, we have

. d daij af A(f) = dt (f(A(t))) = dt axij'

or, since f is arbitrary, .

daij a dt ax']

. a == -(t), dA ax'] dt

A= - .. ==aij..

where summation over repeated indices is understood and we introduced dA/dt as an abbreviation for tiij (a / axij). However, the one-to-one correspondence between matrices and operators makes this more than just an abbreviation. Indeed, we can interpret dA/dt as the derivative of A and perform such differentiation whenever it is possible. The equation above states that

27.1.13. Box. To obtain the matrix elements (coordinates) of the operator

A, one differentiates the t-dependent elements ofthe (matrix representation of the) operator A(t).

27.1 LIE GROUPS ANO THEIR ALGEBRAS

823

Of particular interest are the left iovariant vector fields, or equivalently, the vectors belonging to T,(GL(V)). This amounts to substituting t = 010 the formulas above. Thus, if A E T, (G L (V)),

. a dA A = tiij(O).. es -(0). ax'] dt

(27.8)

For the product of two operators, we get

Ail = !!f(A(t)B(t))[ = dt ,~O

!!-(aikbkj) [ ~ dt 1=0 ax']

.

a

= (tiik(O)bkj(O)+aik(O)bkj(O))-a ., '--v--'

'-.,.-..'

t3kj

a

= tiij(O).. axl]

Differential ofthe determinant map is the trace: det, = tr

(27.9)

x'!

()ik

.

a + bij(O).. = ax']

dA -d (0) t

dB

+ -dt (0).

Many ofthe Lie groups used 10physics are subgroups of G L (V). A characterization of the Lie algebras of these subgroups is essential for understanding the subgroups themselves and applyiog them to physical situations. These subgroups are typically defined 10 terms of maps t/J : G L (V) --> M for which M is a manifold and t/J. is surjective. To construct the Lie algebra of subgroups of GL(V), we need to concentrate on the map t/J" as defined on T, (G L (V)). An important map is det : GL(V) --> iC for a complex vector space V. We are interested 10 evaluating the map det, : T,(GL(V)) --> Tj(iC) 10 which we consider iC ~]R2 to be a manifold. For an operator A E T,(GL(V)) ~ gl(V) and a complex-valued function, we have

. d [ d X af dy af det.(A)f sa - f(detA(t)) es - - + - dt ,~O dt ax dt ay

I

f -af + -d ImdetA(t) [ aax dt 1=0 ay . af . af = RetrA- +ImtrA-, ax ay = -d RedetA(t)

dt

1=0

where we used Equation (3.28). Sioce f is arbitrary and {ajax, ajay} can be identified with (I, i), we have det, (A) = tr A.

(27.10)

27.1.14. Example. LIEALGEBRA OF SL(V) The special linear group SL(V) is characterized by the fact that all its elements have unit determinant.

27.1.15. Box. The Lie algebra sl(V) of the special linear group is the set of all traceless operators.

824

27. LIE GROUPS AND LIE ALGEBRAS

This is becauseif we use (27.10) and (26.12) and thefact that SL(V) = det- l (l), we III can concludethat det.(A) = tr A = 0 for all A E sl(V).

27.1.16. Example. LIE ALGEBRAS OF UNITARY AND RELATED GROUPS Let us first show that the set of unitary operators on V, denotedby U (V). is a Lie subgroup

unilary group

of GL(V), called the unitary group ofV. Considerthe map 1/! : GL(V) --+ 1HI, where IHI is the set of Hermitian operators considered as a vector space (therefore, a manifold) over the reals, defined by 1/!(A) = AAt.UsingEquation (27.9), the readermay verifythat1/!.is surjective and (27.11)

It follows from Theorem 26.3.7 that U(V) _ 1/!-1(1) is a subgroup of GL(V). Using Equation (26.12), we conclude that1/!.(A) = A + At = 0 for all A E u(V), i.e.,

27.1.17. Box. The Lie algebra u(V) of the unitary group is the set of all anti-

hermitian operators.

special unitary group

orthogonal and special orthogonal groups

By counting the number of independent real parameters of a matrix representing a hermitian operator. we can concludethatdim JH[ = n2 . It follows fromTheorem 26.3.7 that dimU(V) =n 2. The intersection of SL(V) and U(V), denotedby SU(V), is called the special unitary group. ItsLie algebra su(V) consistsof anti-hermitian traceless operators. The reader may check that dim SU(V) n 2 - I. When the vector spaceis C", we write U(n) and SU(n) insteadof U(Cn) and SU(Cn). If we restrict ourselves to real vector spaces, then unitary and.special unitary groups becomethe orthogonal group O(V) and special orthogonal group SO(V), respectively. Their algebras consistofantisymmetric andtraceless antisymmetric operators, respectively. When V JRn, we use the ootation O(n) and SO(n). III

=

=

Let X be a vector field OnG. We know from our discussion of flows that X has aflow F, sa exp(tX) at every point g ofG with -E < t < E. Now, since F,(g) i' g is in G, it follows from the group property of G that (F,)n (g) = Fn, (g) E G for all n. This shows that the flow of every vector field on a Lie group is defined for all t E JR, i.e., all vector fields On a Lie group are complete. Now consider 9 as a vector space and manifold and define a map exp : 9 ~ G that is simply the flow evaluated at t = 1. It can be shown that the following result holds ([Warn 83, pp. 103-104]): exponenlial map ofa Lie algebra

27.1.18. Theorem. exp : 9 ~ G, called the exponential map, is a diffeomorphism of a neighborhood of the origin of 9 with a neighborhood of the identity element of G. This theorem states that in a neighborhood ofthe identity element, a Lie group, as a manifold, "looks like" its tangent space there. In particular,

27.1 LIE GROUPS AND THEIR ALGEBRAS

825

27.1.19. Box. Two Lie groups that have identical Lie algebras are locally diffeomorphic.

27.1.20. Example. WHY exp IS CALLEDTHE EXPONENTIAL MAP Let V be a finite-dimensional vector space and A E £1[('\7). Define, as in Chapter 2,

and note that

Furthennore, etAe

sA

00

=

k k

L~

k=O k!

DOGmL =E m=Q

nAn k n E= LL~Ak+n n=O n! k=On=O k!n! 00

t

00

S

00

m-nn) Am = eCt+s)A. S

=0 (m - n)!n! =(t+s)mjm!

It follows that etA has all the properties expected of the flow of the vector field A.

IlIII

The exponential map has some important properties that we shall have occasion to use later. The first of these properties is the content of the following proposition, whose proof is left as an exercise for the reader. 27.1.21. Proposition. Let t/J : H --+ G be a Lie group homomorphism. Then,for all '1/ E ~,we have t/J(exPH '1/) = eXPG(t/J.'1/)·

inner automorphism of a Lie group

For every g E G, let I g es R g1 0 L g • The reader may readily verify that I g, which takes x E G to sxs:" E G, is an isomorphism of G, i.e., Ig(xy) = Ig(x) Ig(y) and Ig is bijective. It is called the inner automorphism associated with g.

adjoint map of a Lie algebra

27.1.22. Definition. The Lie algebra isomorphism I g• = R g 0 L g• : 9 --+ 9 is denoted by Adg and is called the adjoint map associated with g.

!

Using Proposition 27.1.21, we have the following corollary. 27.1.23. Corollary. exp(Adge) = Igexpe = gexpeg-1!orall

G.

e E gandg E

826

27. LIE GROUPS AND LIE ALGEBRAS

Let {ed be a basis for the (finite-dimensional) Lie algebra of the Lie group G. The Lie bracket of two basis vectors, being itself a left-invariant vector field, can be written as a linear combination of (ed:

lei' ej] =

" :~:>tek'

k~!

On a general manifold, c~. will depend on the point at which the fields are being evaluated. However, o~ Lie groups, they are independent of the point, as the following manipulation shows:

n

n

= Lg• I:>~/e)ek(e) k~l

Lie's second theorem

n

= LC~j(e)Lg.ek(e) = Lc~/e)ek(g). k=!

k=!

Therefore, the value of C~j at any point g EGis the sanneas its value at the identity,

i.e., ct is a constant. This statement is called Lie's second theorem. structure constants

27.1.24. Definition. Let G. Then

ofe Lie algebra

(ed?~!

be a basis for the Lie algebra g ofthe Lie group

n

[ei(g), ej(g)] = LC~jek(g),

(27.12)

k=!

0

where j , which are independent of g, are called the structure constants of G.

Lie's third theorem

The structure constants satisfy certain relations that are immediate consequences of the commutation relations. The antisymmetry of the Lie bracket and the Jacobi identity lead directly to

(27.13) The fact that

{~p}

obey Equation (27.l3)is the content of Lie's third theorem.

27.1.3 InfinitesimalAction The action ep : G x M --+ M of a Lie group on a manifold M induces a homomorphism of its algebra with X(M). If e E g, then exp(te) E G can act on M at a point P to produce a curve y(t) = exp(te) . P going through P. The tangentto this curve at P is defined to be the image of this homomorphism. infinitesimal generators ofan action

27.1.25. Definition. Let ep : G x M --+ M be an action. If e E g, then

827

27.1 LIE GROUPS AND THEIR ALGEBRAS

(exptl;, P) is ajlow on M. The corresponding vector field on M given by I;M[P sa I;M(P) sa

~(exptl;, P)[ t=O dt

is called the infinitesimal generator of the action induced by 1;. Infinitesimal generators of representations of G form a representation ofg.

In particular,

27.1.26. Box. If M happens to be a vector space, and the action a represen-

tation as given in Box 24.1.3, then the infinitesimal generators constitute a representation ofthe Lie algebra ofthe group.

27.1.27. Example. Onecan think ofleft transtation on a Liegronp G as an action of G on itself. Let : G x G -+ G be given by (g, h) = Lg(h). Then Definition 27.1.25 gives eG(g)

= ddt (expte, g)1 t=O =!!.exptegl = ddt Rg (expte)1 t=O = Rg.e dt t=O

by thefirst equation in (26.24). It follows that eG is right-invariant. Indeed, eG

0

Rh(g) = eG(gh) = (RghJ.e

Since thishotds for all g

E

= (Rh 0 Rg).e = Rh. 0 Rg.1f. = Rh. 0 If.G(g).

G, it follows that eG 0 Rh = Rh. 0 If.G, demonstrating that If.G

is right-invariant.

adjoint action

11III

The adjoint map of Definition 27.1.22 induces a natural action on the Lie algebra g with some important properties that we now explore. Define the adjoint action : G x g -+ g of G on g ~ ~)"e(G) by (g, 1;) = Adg(I;). We claim that the infinitesimal generator 1;9 is nDe, where nDe(1/) == [1;,1/]. In fact, I;g(1/) = dd (exptl;, 1/)[ t

t=O

= dd Adexpte(1/)[ t

I

1=0

I

d -t d -1 = dt Rexpte. 0 L expte.(1/) t~O = dt Ft. 1/(exptl;) t~O = Le(1/) = [1;,1/] sa nDe(1/),

(27.14)

where we used Equation (27.6) as well as the definition of Lie derivative,Equation (26.31). If : GxM -+ Mis an action, then g : M -+ M,definedbyg(P) = (g, P), is a diffeomorphism of M. Consequently, g. : 'J'p(M) -+ 'J'g.p(M) is an isomorphism for every P E M whose inverse is ct>g; = ep g-l*. 27.1.28. Proposition. Let : G x M -+ M be an action. Then for every g

and 1;,1/

E

g, we have and

E

G

828

27. LIE GROUPS ANO LIE ALGEBRAS

Proof Let P be any point in M. Then. (Adge)M(p) =

~lP(exPtAdge. p)1 dt t=O

=

~lP(g(expte)g-l. p)1

= =

dt

(by Definition 27.1.25) (by Corollary 27.1.23)

t~O

~lP(g(expte). lPrl(pnl t=O

dt

~lPg

dt

= lPg*I

0

lP(expte. lPrl (pnl

g-l

(P)

(by definition of action)

t~O

dd lP(expte. lPrl(pnl

t

(by definition of lPg) (=0

= lPg*Ir1(p) eM (lPrl (P)) = (lPg.eM) (P).

[by (26.24)] [by (26.28)]

The second part of the proposition follows by replacing g with exp t'TJ. so that

Differentiate both sides with respect to t and note that the LHS gives ['TJ. elM. The derivative of the RHS is the Lie derivative of eM with respect to -'TJM. which is -~M.e~.

comparison of coordinate manipulations with geometric (coordinate-free) analysis

D

As mentioned earlier. a Lie group action is usually described in terms of the parameters of the group. which are simply coordinate functions on the group G. as well as coordinate functions on the manifold M. The infinitesimal generators. being vector fields on M. will then be expressed as a linear combination of coordinate frames. In the older literature. no mention of the manifold structure is made. A Lie group is defined in terms of multiplication functions and other functions that represent the action of the group on the manifold. Thus. an r-parameter Lie group G is a

collection of two sets of functions, m!' : IRr ---+ JR, p = 1, 2, ... , r, representing the group multiplication. and cf>i : E' x En -> E. i = I. 2•...• n, representing the action of G on the n-dimensional manifold M. We sketch the procedure below. leaving most of the calculations as exercises for the reader. As we develop the theory. the reader is urged to compare this "coordinate-dependent" procedure with the "geometric" procedure-which does not use coordinates--
x; = tPi(al. .,., ar; Xl, Xi

= tPi(O, ... ,0; Xl,

, Xn),

i = 1, ... ,n,

, Xn ) ,

4Weuse subscripts for coordinate functions herefortypographical convenience.

(27.15)

27.1 LIE GROUPS ANO THEIR ALGEBRAS

829

as well as the group multiplication properties Cp = mp(al,

ap

= mp(O,

, a r;

bI,

, 0; a1,

, br ) ,

P = 1, ... , r,

, ar) = rnp(al. ... , ar; 0, ... ,0),

mp(a; m(b; c)) = mp(m(a; b); c).

(27.16)

Equation (27.15) is to be interpreted as a rule that takes the second set of arguments and transforms them via the first set into the LHS. Now suppose that we translate from x; to a neighboring point x; + dX; via a set of group parameters {~ap}~=l'

We can also get to XI~

+ dX; from Xi

via a new set of parameters.l whichhave to be slightly different from {ap}~=l' say lap + dap}~=l' We then have

xI + dX; = 41i(8al, ... , 8a,; x~, ... , x~), xI + dX; = 4Ji(al

ap + da p =

+ dai, ... , a; + dar; xt , ... , xn),

mp(al, ... , ar;

8a1o ... , 8a,),

(27.17)

and, with summation over repeated indices understood,

dX;

= (27.18)

Inverting the second equation and substituting the resulting ~a's in the first equation yields or where in the last equation, we changed the free coordinate variable on both sides. It then follows that (27.19) Lie's first theorem

Equation (27.19) and establishing that UiK is eoo is the content of Lie's first theorem. The change ofan arbitrary function f(x) due to an infinitesimal transformation is

SHere we areassuming that theactionof thegroup is transitive,i.e., that everypointof themanifold can be connected to any other pointvia a transformation.

830

27. LIE GROUPS AND LIE ALGEBRAS

This suggests calling

a

n

L i=l

XK = infinitesimal generators as vector fields on M

UiK (x)

ij:"

(27.20)

X,

the infinitesimal generators of the Lie group. The commutator of two of these

generators is [Xp,X n] =

aU[o au jp] -a a. [UiP-a--UinaXi Xi Xi

(27.21)

This commutator does not appear to be similar to the one in Definition 27.1.24, which is necessary if the generators are to form a Lie algebra. However, through a long aod tortuous manipulation, outlined in Problem 27.9, one cao show that (27.22)

where c~u are constants. One cao also obtain this sarne result by the much simpler method of applying Proposition 27.1.28 to both sides of Equation (27.12):

[(OM, (ej)Ml = -[e, '7]M = - (tcfjek) k=1

= M

tct(ek)M. k=l

This equation is equivalent to (27.22) if we identify the X p 's with the (ei) M'S aod

ignore the irrelevant minus sign. The reader has hopefully been able to appreciate the power aod elegaoce of the geometric approach to Lie groups aod Lie algebras. The above illustration (Problem 27.9) brings out the tedium aod the error-prone procedure of obtaining group-theoretic results through coordinate maoipulations, a procedure used in the old literature including the work ofSophus Lie himself. Although such calculations areinevitable in practice, wheremostLie groups aregiven in terms of parameters, they are not suitable for obtaining formal results.

27.1.29. Example. The two-dimensional rotation group 80(2) is a I-parameter Lie group defined by

xi == cPl(Xl,XZ; B) =Xl cos s -X2sinB, x2 ee ¢2(XI, xz: 0) = XI sinO + Xz cosO. Using Equation (27.20). we find the (only) generator of this group: X=

Uj

a~i

where Ui

=

aa~ IlbO·

Explicitly, we have "I = "Z

at/>ll

ao

0=0

= (-Xl sinO - Xz cosO)lo=o = -xz,

= at/>zl ao 0=0 =

(XI

cosO -

Xz sinO)lo-o = XI. -

27.1 LIE GROUPS AND THEIR ALGEBRAS

831

and

a a a a a a X=UI-+U2- =-X2-+xI- =-y-+x-. aXI aX2 aXI aX2 ax ay

Thereader recognizes this, within afactorof i, asthez-component of theangular momentum operator. In general,

27.1.30.Box. Angularmomentum operators are the infinitesimal generatorsof rotation.

Inclusion of theother two rotations about thez-axis andthe y-axis completes the set II of infinitesimal generators of therotation group in three dimensions.

The action of a Lie group on M can be reconstructed from its infinitesimal action. The flow of XK is the solution of the DE dx;

dt

, = UiK(X),

X;(O) =Xi.

(27.23)

Once the solution is obtained, one can replace t with UK for each K. In some applications, uiK(x) will be given implicitly in terms of certain parameters of integration of some DEs [unrelated to (27.23)]. The solution of these DEs are typically generators of coordinate transformations that can be wtitten linearly in terms of the parameters. To be more precise, suppose that after solving some DEs, we obtain r

(27.24)

Xi = I>iKfji)(Xl' ... ,Xn), 1C=1

where {eiK] are the parameters of integration, and Xi are components of the vector field that generates the coordinate transformation. This means that for small parameters, one canwrite r

xI = Xi

+ LCiKf;i)(Xl,'."

xn)

K=l

and read off UiK (x) = fji) (x). In that case, we have X[(O) =Xi.

We shall have occasion to use tltis formula later.

(27.25)

832

27. LIE GROUPS ANO LIE ALGEBRAS

27.1.4 Integration on Lie Groups As any other manifold, one can define integration on Lie groups; i.e., one can construct nonvanishing n-forms and use Equation (26.46) to define integrals on a Lie group G. Because of the left-invariant property of objects on G, it would be helpfnl if the integration process were also left-invariant. For this to happen, the n-form would have to be left-invariant. It turns out that this can be accomplished more or less urtiquely: 27.1.31. Proposition. Let G be a Lie group of dimension n. Then there exists a left-invariant nonvanishing n-form 1L that is unique up to a nonzero multiplicative constant. If G is compact, then 110 is also right-invariant and the multiplicative constant can be chosen to be I.

Proof Let lJoe be any nonzero n-form on 'Je(G). The desired n-form is the left translation of this form, i.e., L ;-llJoe' Indeed, let {Xili'=, be left invariant. Then IJog (X, lg, · .. , Xnlg) = L;_llJoe(X,lg,

= lJoe(Lg-l*X,l g ,

, Xnlg) ,

Lg-l*Xn Ig)

= lJoe(Xilg-1g, ... , Xnlg-l g) = lJoe(X,le, ... , Xnle).

This shows that 110 is left-invariant. Now note that any other n-form 1Jo~ on 'Je(G) is a constant multiple of lJoe' Therefore, the corresponding n-form 1Jo~ will be a constant multiple of IJo g • Let x E G and consider 110' '" RilJo. We have

L *gIL ' = L*g

0

R*xl!

=

R*x 0 L*glt

=

R*xJL = JLI •

where we used the fact that L g and R x commute and that 110 is left invariant. The equation above shows that lJo'is also left-invariant. Therefore, 110' = CIJo. If G is compact, we can integrate both sides and note that f G 110 = f G 110' because 110' is D related to 110 by a change of variable. Therefore, c = I and Ri 110 = 110. Haar measure

The left-invariant volume element (nonvanishing n-form) guaranteed by the proposition above is called Haar measure. Since all calculations are done using some coordinate system, we give an explicit expression of the Haar measure in terms of coordinates (parameters) of a general Lie group. Let y = (y', ... , yr) be the coordinates of the translation of x = (x ', ... , x") by g E G. Then we can write y = meg, x), so that dy! = (ayi jaxi)dx i = (ami jaxi)dx'- Therefore,

d y ' A··· A dy r = det (ami(g,X») . dx i A··· A dx r . ax' In particnlar, if x = 0, the coordinates of the identity, then y will be the coordinates of g. So, the volume element at g, denoted by dry, will be given by

d r y = det (ami (~'

ax

X») I

x=o

d rx.

27.2 AN OUTLINE OF LIE ALGEBRA THEORY

833

Note that this is consistent with the geometric definition of the invariant measure given in Proposition 27.1.31 because L.-1* = L g; and the matrix of L; is the inverse of the matrix of L g*. The volume element at g, which is invariant on G'---and therefore has the same value as at the identity-and which we denote by d/1-(g), will be given by

d/1-(g) = d/1-(e) ea d' x = deC! (

density functions associated with Haar measure

ami (g,i

ax

X)) I

d' g,

(27.26)

x=o

where we have replaced y with the more suggestive g. The volume element d' g is the ordinary Euclidean volume element of ll~.' evaluated at the parameters corresponding to g. The quantity multiplying d' g is called the density function. Note that since we are interested in the derivatives of m! at small values of X, we can take the components of X to be small, and retain them only up to the first order. This will sometimes simplify the calculation of the invariant Haar measure. 27.1.32. Example. Fromthemultiplication rulefortheone-dimensional projective gronp given in Example27.1.8, we easily find

Thus the density function is (al a4 - aza3)-2, and the invariant Haar measure is

d/1-(a)

= (a!a4 -

a2a3)-2~a.

ill

27.2 An Outline of Lie Algebra Theory The notion of a Lie algebra has appeared on a number of occasions both in our study of vector fields on manifolds and, more recently, in the study of Lie gronps in the vicinity of their identity elements. Lie algebras play an important role in the representation theory of Lie groups as well. It is therefore worth our effort to spend some time getting acquainted with the formal structure and properties of these algebras. We shall restrict our discussion to finite-dimensional Lie algebras. Lie algebra defined

27.2.1. Definition. A finite-dimensional vector space V over JR (or iC) is called a Lie algebra over JR(or iC) if there is a binary operation, called Lie muUiplication, [".J : V x V --> Von V, satisfying

1. [X, Y] = -[Y, Xl/or all X, Y 2. [aX + fJY, Z] = a[X, Z]

3.

E

V (antisymmetry).

+ fJ[y, ZJ/or a, fJ [X, [Y, Z]] + [Z, [X, Y]] + [Y, [Z, X]] = 0

E

JR (or iC) (linearity). (Jacobi identity).

834

27. LIE GROUPS AND LIE ALGEBRAS

The concepts ofa homomorphism, its kernel, its range, etc. are the same as before.

To distinguish Lie algebras from vector spaces, we shall denote the former by lowercase German letters as we have done for the Lie algebras of Lie groups. 27.2.2. Example. Recall from Chapter t thatanalgebra is a vector space with a product. If this product is associative, then one can construct a Lie algebra out of the associative algebra by defining [a, b] ea ab - ba. In particular, thematrix algebra under commntation of matrices becomes a Liealgebra, which wedenote by gl(n,lit) [or gl(n, iC)]. III

subalgebra, ideal, and center ota Lie algebra

27.2.3. Definition. Let Ube a Lie algebra. A subspace u of» is calleda subalgebra if [X, Y] E u whenever X, Y E u. The subspace u is called an ideal if [X, Y] E.U whenever either X E U or Y E u. The center 3 of U is the collection ofall X E U whose Lie multiplication with all vectors of U vanishes. A Lie algebra is abelian, or commutative, if 3 = n.

Knowing the structure constants,

one can reconstruct the Lie aigebra!

If we choose a basis in the Lie algebra n, and express the Lie multiplication of basis vectors as a linear combination of basis vectors, we end up with basisdependent structureconstantsthat satisfyEquation (27.13).The structnreconstants completely determine the Lie algebra: Given these constants, one can choose a vector space V of correct dimension, a basis in that space, and impose the Lie multiplication law among the basis vectors suggested by the structure constants. Once the Lie multiplicationlaw for basisvectors is established,the law for arbitrary vectors follows from linearity of Lie multiplication. This procedure induces a binary operation on V and turns it into a Lie algebra n. Any other algebra so constructed will be isomorphic to n,

27.2.4. Example. We can classify all two-dimensional Lie algebras by analyzing their structure constants. Let Xl and X2 be any two linearly independent vectors of the twodimensional Lie algebra n, Write the only nonzero Lie bracket as [Xl, X2] = CjXj

+ C2X2·

There are twocases to consider: Either C1 = 0 = C2 orone of theconstants is nonzero. The first case corresponds to a 2-dimensional abelian Lie algebra: fori,j

= 1,2.

Forthe secondcase, suppose C1 ':f::. 0 anddefine the vectors Y es X2/Cj.

Thenthenonzero Lie bracket becomes [X, Y] = X.

The result of Example 27.2.4 is summarized as follows: 27.2.5. Box. There are only two 2-dimensional Lie algebras given by either one of the following nonzero Lie bracket relations: or

iii

27.2 AN OUTLINE OF LIE ALGEBRA THEORY

835

27.2.6. Example. The Pauli spiu matrices

cz

>

0 (i

-i) 0 '

fcrm a Lie algebra under the commutation relation given by

Thus, c~k = 2iEiki- Pauli spin matrices are a basis for su(2).

Wayl basis for gl(n,lll.)

III

27.2.7. Example. The Lie group GL(n, lll.) hasgl(n, R), the set of all real n xn matrices, as itsLie algebra.The standardbasisof thisLie algebra,also calledthe Weylbasis, consists of matrices 8ij that have zeros everywhere except at the ijth position. We therefore have (27.27) We can readily find the Lie multiplication (commutation relations) for these matrices. We simply need to look at the elements of the matrix of the commutator: ([Sij, eklD mn

= (8ijekl)mn - (ekleij}mn = (eij)mr (ekllrn - (ekl)mr (eij),n =~~r~~-~~~r~n=~m~~-~~~n

= (eU}mn 8jk -

(ekj)mn 8Jj.

or (27.28) The structure constants, which are naturally double-indexed,can be read off from Equation (27.28): (27.29) where we have used a superscript for some of the Kronecker deltas to conform to the position of the corresponding index on the LHS. III 27.2.8. Example. An important datum is the dimension of the Lie group (or its associated Lie algebra, since they are the same). This datum is not apparent in most cases of interest in which the group is defined in terms of some geometric property. Forexample, the symplectic group is defined as all linear transfonnations A that leave a certain antisymmetric bilinear fonn invariant (Example 23.2.2). In tenns of matrices, we have

J=(O

-1

It follows that the symplectic group consists of all matrices A such that

A'JA=J.

(27.30)

836

27. LIE GROUPS AND LIE ALGEBRAS

If we write A in block form, A = (At!

\.A21

whereAij are n x n matrices, then,Equation (27.30) becomes 1) (At!

o

A21

A12) A22

=

( 0 -1

~) ,

or (27.31)

Forthe symplecticalgebra sp(2n,lR.), we areinterested in the matrix Awhen it is close to theidentity. Thismeans that

Substituting these in Equation (27.31) and keeping terms linear in E, we obtain the following

relations among Xij: (27.32)

It follows that we need n 2 parameters to describe the n x n matrices XII andX22 simultaneously. For the symmetric matrices X12 andX2I. we need n(n + 1)/2 independent parameters

each.Therefore, thetotalnumber of independent parameters neededfor (orthe dimension of) the symplectic algebra sp(2n,lR) is

III

Although our attempt is to give a formal discussion of the Lie algebras and their structure in this section, we shall do this with an eye to the eventual utility of this discussion in a better understanding of the Lie algebras of Lie groups. To make the connection between the present formalism and the Lie algebras arising from Lie groups, we shall make heavy use of matrix groups, i.e., GL(n,IR) [or GL(n, iC)] and its subgroups. Equation (27.8) gives a method of finding the matrices of the algebra if those of the group are known:

27.2.9. Box. Differentiate the matrix with respect to a parameter at the identity (where all parameters are set equal to zero) to find the matrix "in the direction" of that parameter.

27.2 AN OUTLINE OF LIE ALGEBRA THEORY

837

27.2.1 The Lie Algebras o(p,n - p) and p(p,n - p) Many of the Lie groups encountered in physical applications are special cases of the (pseudo) orthogonal group 0 (p, n - p) and its associated Poincare group p (p, n - p). It is therefore worthwhile to study their Lie algebras in some detail. Introduce the diagonal matrix .., = diag(-I, -I, ... , -1, 1, 1, ... ,1)

.

''-.,.--'

p times

n - p times

and note that the (pseudo) orthogonal group 0 (p, n - p) consists of n x n matrices that leave the bilinear form x . x == x'..,x invariant for x E En. This means that the matrices A will have to satisfy A'..,A worthogonal matrices

=..,

=}

(detA)2

=1.

(27.33)

Such matrices are called ..,-orthogonal. The fact that 0 (p, n - p) is a group and that = .., can be used to show that

..,-1

(27.34) 27.2.10. Example. TIlE LORENTZ GROUP The group of the special theory of relativity is the full Lorentz group 0(3, 1). This is the group of transformations that leave the invariant length6 T/ijXiXj =

-xr -xi -

x~

+xt ee x6 -xr -xi - xi

of a 4-vector (XI, X2, X3, Xo = ct) iovatiant. The (0, OJ-components of Equations (27.33) and (27.34) yield 2 - alQ 2 aOO

2 - a 2 = 1, llZo 30

a50 - a51 - a52 - a53 = 1. ormoehronous and proper ortnocnronous Lorentz transformations Timelike, spacelike, and null vectors: JR' as the set of events

(27.35)

Either one of these equations implies that aOO 2: 1 or aOO ~ -1. Lorentz transformations for which coo ~ 1 are called orthuchronons. Since detl = +1 and 100 = +1, the identity belongs to the subset consisting of transformations with det A

== +1 and aoo

::: 1. Such

transformations fonn a snbgroup of 0(3,1) called the proper urthochronuns Lorentz transformations, and have the property that they can be reached continuously from the identity. Depending on whether x . x > 0, x . x < 0, or x . x = 0, the vector x is called timelike, spacelike, or null, respectively. In the special theory of relativity IR4 becomes the set of events. At every event x the set ]R4 is divided into 5 regions:

I. All eventsy = (Ylo Y2, Y3, YO) to which one can go from x by material objects,with speed less than c.Iie to the future of x, i.e., YO (YO -xO)2> (YI _XI)2

future light Gone

xo c- 0, and are timelike:

+ (Y2 -X2)2 + (Y3 -X3)2.

They form a 4-dimensional subset of IR4 and are said to lie inside the future light cone.

6n is common to label the time coordinate with index 0 rather than 4. We shall use this convention.

838

27. LIE GROUPS ANO LIE ALGEBRAS

2. All events Y= (YI, Y2, Y3, YO) to which one can go from x only hy a light signal lie to the future of x, i.e., YO - xo > 0, and (YO - xO)2 - (YI - XI)2 - (Y2 - X2)2 - (Y3 - X3)2

= o.

Theyform a 3-dimensional subsetof]R.4 andaresaid to lie on the future lightcone. 3. All events y = (Yl, Y2, Y3, YO) from which one can come to x by material objects, with speedless than c. lie in thepastof x, i.e., XQ - YO > 0, andaretimelike: (XO - YO)2 > (xi - YI)2

+ (X2 -

Y2)2

+ (X3 -

Y3)2.

They form a 4-dimensional subset of lR4 and are said to lie inside the past light cone.

past light cone

4. All events y = (Ylo Y2, Y3, YO) from which one can come from x only hy a light signal lie to the past of x, i.e.,

XQ -

YO > 0, and

(xo - YO)2 - (Xl - YI)2 - (X2 - Y2)2 - (X3 - Y3)2 = O.

They fonn a 3-dimensional subset of Jll.4 and are said to lie on the past light coue. 5. All events in the remaining part of lR.4 form a 4-dimensional subset, are spacelike,

andcannot be connected to x by anymeans. They aresaid to belong to elsewhere.

elsewhere

From a physicalstandpoint, future andpast are observer-independent. Therefore, if y lies in or on the future light cone of x with respect to one observer, it shouldalso do so withrespect to all observers. Since observers areconnected by Lorentz transformations, we expectthe latter to preserve this relation betweenx and y. Not all elements of 0(3,1) have this property. However, the proper orthochronous transformations do. The details are left III as a problem for the reader (see Problem 27.14). As a prototype of "I-orthogonal matrices, consider the matrix obtained from the unit matrix by removing the iith, ijth, jith, and jjth elements, and replacing them by an overall 2 x 2 matrix. The result, denoted byA (ij), will look like

A(ij) =

I 0 0 I

0 0

0 0

0 0

0 0

au

aij

0

0 0

an

ajj

0

0 0

0

0

I

This matrix will traasform (Xl, ... ,Xn ) E

,

Xi = aiiXi +aijXj, t

Xj = ajiXi +ajjXj,

,

Xk =Xk

forkt=i,j.

~n

(no summation!)

according to

27.2 AN OUTLINE OF LIE ALGEBRA THEORY

839

In order for A(ij) to leave the bilinear form X'T/X invariant, the 2 x 2 submatrix (:~~ ;~~) must be either a rotation (corresponding to the case where i, j :::: p or i, j > p), or a Lorentz boost? (corresponding to the case where i ::; P and j > p). In the first case, we have

- Sine) CDSe

'

and in the second case - sinh

g)

coshs

,

where g ea tanh-I (v/c) is the "rapidity." The matrices of the algebra are obtained by differentiation ate = 0 (or g = 0). Denoting these matrices by Mij, we readily find that for the case of rotations, Mij has -I at the ijth position, +I at the jith position, and 0 everywhere else. For the case of boosts, Mij has -I at the ijth and the jith position, and 0 everywhere else. Both cases can be described by the single relation .. )m1 -_ TJilUj "" (MI)

-1}jl o,m i I

Mij

= -Mji.

It is convenient to have all indices in the lower position. So, we multiply both sides by rymk to obtain (27.36)

Lie brackets forthe algebra alp,n - p)

We can use Equation (27.36) to find the Lie multiplication (in this case, matrix commutation relations) for the algebra 0 (p, n - p): (27.37)

Poincare group as

The Lie group 0 (p, n - p) includes rotations and Lorentz transformation. Another group with considerable significance in physics is the Poincare group p (p, n - p), which includes translations'' in space and time as well. An element of pep, n - p) transforms x E JR" to x' = Ax + u, where U is a columo vector representing the translation part of the group. It is convenient to introduce matrices to represent these group operations. This is possible if we represent an element of JR" as an (n + I)-column whose last element is an insignificant I. then, the reader may check that a Poincare transformation can be written as

(n+1) x (n+1)

matrices

(27.38)

7Theelementary Lorentz transfonnationsinvolving only one space dimension. 80ne can thinkof the Poincare group as a subgroup of the group of affine motionsin which thematrices belongto 0 (p, n - p) rather than GL(n, JR.).

840

27. LIE GROUPS ANO LIE ALGEBRAS

where A is the n x n matrix of O(p, n - p), and u is an n-dimensional column. vector. The Lie algebra of the Poincare group is obtained by differentiating the (n + I) x (n + 1) matrix of Equation (27.38). The differentiation of the matrix A will give o(p, n - p) of Equation (27.37). The translation part will lead to matrices Pi with matrix elements given by

(pdz = 8f8

z+! '*

(Pi)kZ = ~ik8Z+!·

(27.39)

These matrices satisfy the following Lie multiplication rules:

It then follows that the full Poincare algebra p (p, n - p) is described by the Lie brackets forthe Poincare algebra p(p, n - p)

following Lie brackets: [Mij, MkzJ = ~ikMjZ - ~ilMjk

+ ~jlMik -

~jkMil,

[Mij, Pk] = ~ikPj - ~jkPi, [Pi, Pj] =

o.

(27.40)

27.2.2 Operations on Lie Algebras 27.2.11. Definition. Let u beaLie algebra. A linearoperator 0 : U --> Usatisfying O[X, Y] = [OX, Y]

+ [X, OY]

is calleda derivation of n, derivation algebra of a Lie algebra

adjoint algebra of a Lie aigebra

Illustration of homomorphism of su(2) and its adjoint using Pauli spin

matrices

Although the product of two derivations is not a derivation, their commutator is. Therefore, the set ofderivations of a Lie algebra Uthemselves form a Lie algebra :D. under commutations, which is called the derivation algebra. Recall that the infinitesimal generators of the adjoint action of a Lie group on its algebra were given by a~e [Equation (27.14)]. We can apply this to a general Lie algebra U by fixing a vector X E Uand defining the map adx : U --> U given by adx(Y) = [X, Y]. The reader may verify that adx is a derivation of Uand that ad[X,Y] = [OOx, adv], Therefore, the set a~. == {OOx I X E n] is a Lie algebra, a subalgebra of the derivation algebra D; of U, and is called the adjoint algebra of U. There is a natoral homomorphism 1/r : U --> a~. given by 1/r(X) = adx whose kernel is the center of e. Furthermore, a~. is an ideal of :D e-

27.2.12. Example. We construct the matrix representation of the operators in the adjoint algebra of 5u(2) with Pauli spin matrices as a basis. From

we conclude that the first column of the matrix of ad a l is zero. From

841

27.2 AN OUTLINE OF LIE ALGEBRA THEORY

we conclude that thesecondcolumn of thematrix of ada1 haszerosforthefirst twoelements and 2i for the last. Similarly, from

we conclude that the third columnof the matrix of adul has zeros for the first andthird elements and- 2i forthe second. Thus, thematrix representation of adrr1 is 0 adot = 0 (

o

0 0

2i

Likewise,we can obtain the othertwo matrix representations; theyare

adu2=(~-2i

oo o

2i) 0

-2i

o

,

o

0

~) .

The reader may readily verify that [adaj , adu k ] = 2iEjkladq.

III

If 1jf is an automorphism of n, i.e., an isomorphism of n onto itself, then

ady,(X) =

1/J a adx a 1/J-l

VXE~.

(27.41)

Since adx is an operator on the vector space e, one can define the trace of adg , However, the notion of trace attains a far greater sigoificance when it is combined with the notion of composition of operators. For X, Y E e, define

(X [Y) sa tr(adx

0

(27.42)

ady).

Then one can show that (.[.) is bilinear and symmetric. It becomes an inner product if the "vectors" of the Lie algebra are hermitian operators on some vector space V, or if the underlying vector space is over ~ (see Proposition 3.6.6 in Chapter 3). Furthenmore, (.[.) satisfies

([X, Yll Z) + ([X, Z] IY) = O. Killing lorm 01 a Lie algebra

27.2.13. Definition. The symmetric bilinear form (.[.) : (27.42) is called the Killing form of»,

(27.43) ~

x

~

--> n defined by

It is an inunediate consequence of this definition and Equation (27.41) that the Killing fonm of ~ is invariant under all automorphisms of n.

Most mathematicians seem to have little orno interestin history, so that often thename attached to a key result is that of the follow-up person who exploited an idea or theorem rather than its originator (the Jordan form is dueto Weierstrass, Wedderburn theory to Cartan and Malien). No one. hassuffered fromthisahistoricism morethan Killing.Forexample, the so-called"Carlan sub-algebra"and "Carlan matrix" were definedand exploitedby Killing.

842

27. LIE GROUPS AND LIE ALGEBRAS

He exhibited the characteristicequation of an arbitrary element of the Weyl group when Weyl was 3 years old and listed the orders of the Coxeter transformation 19 years before Coxeter was bern! Wilhelm Karl Joseph Killing (1847-1923) began university study in Munster in 1865 but quickly moved to Berlin and came under the influence of Kummer and Weierstrass. From 1868 to 1882 much of Killing's energy was devoted to teaching at the gymnasium

level in Berlin and Brilon (south of Miinster). At one stage, when Weierstrass was urging him to write up his research on space structures, he was spending as much as 36 hours per week in the classroom or tutoring. (Now many mathematicians consider 6 hours a week an intolerable burden!) On the recommendation ofWeierstrass, Killing was appointed professor of mathematics at the Lyzeum Hosianum in Braunsberg, in East Prussia (now Braniewo in the region of OIsztyn, in Poland). This was a college fouuded in 1565by Bishop Stanislaus Hosius, whose treatise on the Christian faithran to 39 editions! The main object ofthe college was the training of Roman Catholic clergy, so Killing had to teach a wide range of topics, including the reconciliation of faith and science. Although he was isolated mathematically during his ten years in Braunsberg, this was the most creative period in his mathematical life. Killing produced his brilliant work despite worries about the health..!of his wife and seven children, demanding administrative duties as rector of the college and as a member and chairman of the City Council, and his active role in the Church of St. Catherine. What we now call Lie algebras were invented by the Norwegian mathematician SOphU5 Lie about 1870 and independently by Killing about 1880. Lie was seeking to develop an approach to the solution of differential equations analogous to the Galois theory of algebraic equations. Killing's consuming passion was non-Euclidean geometries and their generalizations, so he was led to the problem of classifying infinitesimal motions of a rigid body in any type of space (or Raumformen, as he called them). In 1892 he was called back to his native Westphalia asprofessor of mathematics at the University of Miinster, where he was quickly submerged in teaching, administration, and charitable activities. He was Rector Magnificus for some period and president of the St. Vincent de Paul charitable society for ten years. Killing's work was neglected partly because he was a modest man with high standards; he vastly underrated his own achievement. His interest was geometry, and for this he needed all real Lie algebras. To obtain merely the simple Lie algebras over the complex numbers did not appear to him to be very significant. Another reason was due to Lie, who was quite negative about Killing's work. At the top of page 770 of a three-volume joint work of Lie and Engel we find the following less than generous comment about Killing: "With the exception of the preceding unproved theorem. .. all the theorems that are correct are due to Lie and all the false ones are due to Killing!" Killing was conservative in his political views and vigorously opposed the attempt to reform the examination requirements for graduate students at the University of Miinster by deleting the compulsory study of philosophy. Engel conunents "Killing could uot see that for most candidates the test in philosophy was completely worthless," He had a profound patriotic love of his country, so that his last years (1918-1923) were deeply pained by the collapse of social cohesion in Germany after the War of 1914-18.

27.2 AN OUTLINE OF LIE ALGEBRA THEORY

843

(Taken from A. J. Coleman, "The Greatest Mathematical Paper of All Times," Mathematical Intelligencer 11 (3) (1989) 29-38.)

As noted above, the Killing form is an inner product if the Lie algebra consists of hermitian operators. This will certainly happen if the Lie algebra is that of a group whose elements are unitary operators on some vector space V. We shall see shortly that such unitary operators are not only possible, but have extremely useful properties in the representation of compact Lie groups. A unitary representation of a Lie group induces a representation of its Lie algebra whose "vectors'; are hermitian operators. Then the Killing form becomes an inner product. The natnral existence of such Killing forms for the representation of a compact Lie group motivates the following: compact Lie algebra

27.2.14. Definition. A Lie algebra U is compact satisfying

([X, Y] 1 Z)

+ ([X, Z]

1

if it

has an inner product (·1·)

Y) = O.

Choose a basis {Xi} for the Lie algebra U and note that (adx,)j = cfj' Therefore, (27.44) Cartan metric tensor ofa Lie algebra

where gij are components of the so-called Cartan metric tensor in the basis {Xii. If A, B E U have components {ai} and {bi} in the basis {Xi}, then it follows from Equation (27.44) that

j

(27.45)

(A 1 B) = aib gij,

as expected of a symmetric bilinear form. We can use the Cartan metric to lower the upper index of the structnre constants: Cijk es C!jglk. By virtue of Equation (27.44), the new constants may be written in the form by (27.13)

The reader may now verify that the RHS is completely anti symmetric in i, j, and k. If the Lie algebra is compact, then one can choose an orthonormal basis in which glk = 81k (because the inner product is, by definition, positive definite) and obtain cfj = Cijk- We therefore have the following resnlt. 27.2.15. Proposition. Let U be a compact Lie algebra. Then there exists a basis of U in which the structure constants are represented by a third-order completely antisymmetric covariant tensor.

844

27. LIE GROUPS AND LIE ALGEBRAS

27.2.16. Example. We can calculate explicitly the Killing form of the Lie algebras glen, lR) and olen, R). Choose the Weyl basis introduced in Example 27.2.7 and expand A, B E g(n, lR) in terms of the Weyl basis vectors: A = aij eij. B = hi} ef}. The Cartan metric tensor becomes gij,kl = ciJ.mnck}~rs = (~jm8[ 8~ -

8in8:-ntSj) (81 r tSr 8:

:'8i ),

- 8ks 8

where we have used Equation (27.29). It follows from these relations, Equation (27.45), and a simple index manipulation that (A I B) == a ij bkl gij,kl = 2n tr(AB) - 2 tr Atr B

(27.46)

for A,B E glen, R), and (A I B) = 2n tr(AB)

(27.47)

for A, B E sl(n, R), because all matrices in sl(n, JR.) are traceless.

III

A Lie algebra n, as a vector space, may be written as a direct sum of its snbspaces. We express this as r

U

= UI +

U2 + ... + Ur

=L

+Uk.

k=l

If in addition {Uk} are Lie subalgebras every one of which commutes with the rest, we write r

U = UI Ell U2 Ell ... Ell Ur

= L EIluk

(27.48)

k=l

and say that Uhas been decomposed into a direct sum ofLie algebras. In this case, each Uk is not only a subalgebra, but also an ideal of Uas the reader may verify. The study of the structure of Lie algebras boils down to the study of the "simplest" kind ofLie algebras in terms of which other Lie algebras can be decomposed. Intuitively, one would want to call a Lie algebra "simple" if it has no proper subalgebras. However, in terms of decomposition, such subalgebras are required to be ideals. So the natural definition of a simple Lie algebra would be the following: semisimple Lie algebras

27.2.17. Definition. A Lie algebra that has no proper ideal is called a simple Lie algebra. A Lie algebra is semisimple ifit has no (nonzero) commutative ideal. For example, the pseudo-orthogonal algebra o(p, n - p) is setuisimple, but the Poincare algebra p (p, n - p) is not because the translation generators P j form a commutative ideal. A useful criterion for semisimplicity is given by the following theorem due to Carlan, which we state without proof (for a proof, see [Baru 86, pp. 15-16]): 27.2.18. Theorem. (Carlan) A Lie algebra Uis semisimple ifand only ifdet (gij)

o.

oF

27.3 REPRESENTATION OF COMPACT LIE GROUPS

845

The importance of semisimple Lie algebras is embodied [Barn 86, pp. 19-20]. 27.2.19. Theorem. (Carlan) A semisimple complex or real Lie algebra can be decomposed into a direct sum of pairwise orthogonal simple subalgebras. This decomposition is unique up to ordering. The orthogonality is with respect to the Killing form. Theorem 27.2.19 reduces the study of semisimple Lie algebras tn that of simple Lie algebras. What about a general Lie algebra U? If U is compact, then it turns out that it can he written as U = 3 Ell5 where 3 is the center of U and 5 is semisimple. If U is not compact, then the decomposition will not be in terms of a direct sum, but in terms of what is called a semidirect sum one of whose factors is semisimple. For details, the reader is referred to the fairly accessible treatment of Barut and Raczka, Chapter I. From now on we shall restrict our discussion to semisimple Lie algebras. These algebras are completely known, because simple algebras have been completely classified. We shall not pursue the classification of Lie algebras. However, we simply state a definition that is used in such a classification, because we shall have an occasion to use it in the representation theory of Lie algebras. Cartan subalgebra andthe rank of a Lie algebra

27.2.20. Definition. Let Ube a Lie algebra. A subalgebra ~ of» is called a Cartan subalgebra if~ is the largest commutative subalgebra of », and for all X E ~,if adx leaves a subspace of» invariant, then it leaves the complement of» invariant as well. The dimension of~ is called the rank of».

27.3 Representation of Compact Lie Groups Representation of general Lie groups is closely related to representation of their Lie algebras, and we shall discuss them in the next two sections. In this section, however, we shall consider the representation of compact Lie groups, because for such groups, many of the ideas developed for finite groups hold. Before discussing compact groups, let us state a definition and a proposition that hold for all Lie groups. 27.3.1. Definition. A representation of a Lie group G an a Hilbert space 1f is a Lie group homomorphism T : G --> GL(1f). Similarly, a representation of the Lie algebra 9 is a Lie algebra homomorphism X : 9 --> g[(1f). The proposition we have in mind is the important Schur's lemma which we state without proof (for a proof see [Barn 86, pp. 143-144]). Schur's lemma

27.3.2. Proposition. (Schur's lemma) A unitary representation T : G --> GL(1f) of a Lie group G is irreducible if and only if the only operators commuting with all the Tg are scalar multiples ofthe unit operator.

846

27. LIE GROUPS ANO LIE ALGEBRAS

27.3.3. Example. COMPACTNESS OF U(n), O(n), 8U(n), AND 80(n) 2

Identify GL(n, C) with ]R2n via components. The map / : GL(n,
given by

is continuous because it is simply the products of elements of matrices. It follows that , 2 /-1 (1) is closed,because thematrix 1 is a singlepoint. in ]R2n • whichis therefore closed. /-1(1) is also bounded, because n

L aijakj

j=l

n

L

= fJik =} laij . i,j=l

2 1

= n.

2

Thus, /-1(1) is a (2n 2 -I)-dimensional sphere of radius ,;n in ]R2n , which is clearly bounded. The BWHB theorem (of Chapter 16) now implies that /-1(1) is compact. Now note that 1-1 (1) consists of all matrices thathave theirhermitian adjoints for an inverse; buttheseareprecisely the set U (n) of unitary matrices. Now consider the map det : U(n) -+ C. This map is also continuous, implying that det'" 1(I) is a closed subset of U (n). The boundedness of U (n) implies that det'" 1(I) is also bounded. Invoking the BWHB theorem again, we conclude that det'" 1(I) = 8U(n), being closed andbounded, is compact. If instead of complex numbers, we restrict ourselves to thereals, 0 (n) andSO (n) will replace U(n) and8U(n),respectively. III

The result of the example above can be summarized:

27.3.4. Box. The unitary U(n), orthogonal O(n), special unitary SU(n), and special orthogonal SO(n) groups are all compact.

All representations of compact groups can be made unitary.

We now start our study of the representations of compact Lie groups. We first show that we can always assume that the representation is unitary,

27.3.5. Theorem. Let T : G --> GL(X) be any representation of the compact group G. There exists a new inner product in 1C relative to which T is unitary.

Proof Let ( I) be the initial inner product. Define a new inner product ( I) by

27.3 REPRESENTATION OF COMPACT LIE GROUPS

847

where d /-'g is the Haar measure, which is both left-and right-invariant. The reader may check that this is iudeed an inuer product. For every h E G, we have

(ThuIThV) = fa (TgThUITgThV)dl-'g = fa (TghUITghV)dl-'g

(because T is a representation)

= fa (TghUITghV)dl-'gh

(because I-'g is right invatiant)

= (a]u). This shows that This unitary for all h

E

D

G.

From nowon,we shallrestrict ourdiscussion tounitary representations of compact groups. The study of representations of compact groups is facilitated by the following

construction:

Weyl operator fora compact Lie group

27.3.6. Definition. Let T : G --> GL(J{) be a unitary representation of the compact group G and Iu) E X a fixed vector. The Weyl operator Ku associated with'lu) is defined as

Ku = faiTgU) (Tgul dl-'g.

(27.49)

The essential properties of the Weyl operator are summatized in the following:

27.3.7. Proposition. Let T : G --> GL(X) be a unitary representation of the compact group G. Then the Weyl operator has the following properties 1. Ku is hermitian. 2. KuTg = Tg Ku for all g E G. Therefore, any eigenspace ofKu is an invariant subspace ofall Tg' s.

3. Ku is a Hilbert-Schmidt operator.

Proof Statement (I), in the form (wi Ku Iv)' = (vi Ku Iw), follows directly from the definition. (2) From r, fa ITxu) (Txul dl-'x = fa ITg Txu) (Txul dl-'x' the fact that T is a representation (therefore, T gTx = Tgx), and redefining the integration vatiable to y = gx ; we get TgKu =

r

Ja

ITyu) (Tg-1yul dl-'g-ly =

~ =dJLy

r

Ja

ITyu) (Tg-1Tyul dl-'y,

848

27. LIE GROUPS AND LIE ALGEBRAS

where we used the left invariance of f.L and the fact that T is a representation. Unitarity of T now gives TgKu =

L

ITyu) (T;Tyul df.Ly =

L

ITyu) (TyuITgdf.Ly = KuTg.

(3) Recall that an operator A E £.(J£) is Hilbert-Schmidt if L~l IIA lei) 11 2 is finite for any orthonormal basis {lei}} of J£. In the present case, we have Ku lei) =

L

ITxu) (Txul ei) df.Lx·

Therefore,

f:

2

f: (1 = f: r r Ja Ja

IIKu lei} 11 =

i=l

(eil Tyu) (Tyul df.L y) (lITxU) (Txul ei) df.Lx)

G

i=l

G

(eil Tyu) (Tyul Txu) (Txul et) df.Lx df.L y.

i=l

If we switch the order of summation and integration and use 00

L (Txul ei) (eil Tyu) = (Txul Tyu), i=l

we obtain

and using the Schwarz inequality in the integral yields

f:

IIKu lei) 11

2

:s:

i=l

=

rr

Ja Ja

LL

= lIull 4

(Txul Txu) (Tyul Tyu) df.Lx df.L y (ulu) (ulu}df.Lxdf.Ly

LL df.Lx

where YG is the finite volume of G.

df.Ly

= lI ull 4

(because rep. is unitary)

Ya<

00,

D

Hermann Klaus Hugo Weyl (1885-1955) attended the gymnasium at Altonaand, on the recommendation of theheadmaster afhis gymnasium, who was a cousinof Hilbert, decided attheage of eighteento enter theUniversity of Gottingen. Exceptforone yearatMunich he

27.3 REPRESENTATION OF COMPACT LIE GROUPS

849

remained at GOttingen, as a student and later as Privatdozent, until 1913. when he became professor at the Universityof Zurich. After Klein's retirement in 1913,Weyl declined an offer to be his successor at Gottingen but accepted a second offer in 1930, after Hilberthad retired. In 1933 he decided he could no longer remain in Nazi Gennany and accepted a position at the Institute for Advanced Study at Princeton, where he worked until his retirement in 1951. In the last years of his life he divided his time between Zurich and Princeton. Weyl undoubtedly was the most gifted of Hilbert's students. Hilbert's thought dominated the first part of his math-

ematical career; and although later he sharply diverged from his master, particularly on questions related to foundations of mathematics, weyl always shared his convictions that the value of abstract theories lies in their success in solving classical problems and that the proper way to approach a question is through a deep analysis of the concepts it involves rather than by blind computations. Weyl arrived at Gottlngen during the period when Hilbert was creating the spectral theory of self-adjoint operators, and spectral theory and harmonic analysis were centralin Weyl's mathematical research throughout his life. Very soon, however, he considerably broadened the range of his interests, including areas of mathematics into which Hilbert had never penetrated, such as the theory of Lie groups and the analytic theory of numbers, thereby becoming one of the most universal mathematicians of his generation. He also had an important role in the development of mathematical physics, the field to which his most famous books, Raum, Zeit und Materie . (1918), on the theory of relativity, and Gruppentheorie und Quantenmechanik (1928), are devoted. Weyl's versatility is illustrated in a particularly striking way by the fact that immediately after some original advances in number theory (which he obtained in 1914), he spent more than ten years as a geometer-a geometer in the most modem sense of the word, uniting in his methods topology, algebra, analysis, and geometry in a display of dazzling virtuosity and uncommon depth reminiscent of Riemann. Drawn by war mobilization into the German army, Weyl did not resume his interrupted work when he was allowed to return to civilian life in 1916. At Zurich he had worked with Einsteinfor one year, and he became keenly interested in the general theory of relativity, which had just been published; with his characteristic enthusiasm he devoted most of the next five years to exploring the mathematical framework of the theory. In these investigations Weyl introduced the concept of what is now called a linear connection, linked not to the Lorentz group of orthogonal transformations, but to the enlarged group of conformal transformations; he even thought for a time that this would give him a unified theory of gravitation and electromagnetism, the forerunner of what is now called gauge theories. Weyl's use of tensor calculus in his work on relativity led him to reexamine the basic methods of that calculus and, more generally, of classical invariant theory that had been its forerunner but had fallen into near oblivion after Hilbert's work of 1890. On the other hand, his semiphilosophical, semimathematical ideas on the general concept of"space" in connection with Einstein's theory had directed his investigations to generalizations of Helmholtz's problem of characterizing Euclidean geometry by properties of "free mobility." From these two directions Weyl was brought into contact with the theory of linear representations of

850

27. LIE GROUPS AND LIE ALGEBRAS

Lie groups; his papers on the subject (1925-1927) certainly represent his masterpiece and must be counted among the most influential works in twentieth-century mathematics. Based on the early 19008 works of Frobenius, I. Schur, and A. Young,Weyl inaugurated a new approach for the representation of continuous groups by focusing his attention on Lie groups, rather than Lie algebras.

VeryfewofWeyl's 150publishedbooksandpapers-eventhosechieflyof anexpository character-lack an original idea or a fresh viewpoint. The influence of his works and of his teaching was considerable: He proved by his example that an "abstract" approach to mathematics is perfectly compatible with "hard" analysis and, in fact, can be one of the most powerful tools when properly applied. Weyl was one of that rare breed of modem mathematician whose contribution to physics was also substantial. In an interview with a reporter in 1929, Dirac is asked the following question: " . .. I want to ask you something more: They tell me that you and Einstein are the only two real sure-enough high-brows and the only ones who can really understand each other. I won't ask you if this is straight stuff, forI know you are too modest to admit it. But I want to know this-Do you ever run across a fellow that even you can't understand?" To this Dirac replies one word: "Weyl." Weyl had a lifelong interest in philosophy and metaphysics, and his mathematical activity was seldom free from philosophical undertones or afterthoughts. At the height of the controversy over the foundations of mathematics, between the formalist school of Hilbert and the intuitionist school of Brouwer, he actively fought on Brouwer's side. His own comment, stated somewhat jokingly, sums up his personality: "My work always tried to unite the truth with the beautiful, but when I had to choose one or the other, I usually chose the beautiful."

We now come to the most fundamental theorem of representation theory of compact Lie groups. Before stating and proving this theorem, we need the following lemma: 27.3.8. Lemma. Let T : G ...... G L (1f) be an irreducible unitary representation ofa compact Lie group G. For any nonzero lu) , Iv) E 1f, we have (27.50)

where c > 0 is a constant independent oflu) and [u). Proof. By Schur's lemma and (2) of Proposition 27.3.7, Ku = ).(u)1. Therefore, on the one hand, (27.51) On the other hand,

(27.52)

27.3 REPRESENTATION OF COMPACT LIE GROUPS

851

Moreover, if we use a,.. = a"g-l (see Problem 27.11), theu

fa (vi r, lu) (ul Tll v) a,.x = fa (vi r, lu) (ul T;l Iv) a,.x = fa (uITx-llv) (vlT x lu) a,.x = fa (ulTy Iv) (vITy-llu) a"y-l

(vi Ku Iv) =

= { (ul Ty [u) (vi rt lu) a"y-l = (ul Kv lu). Ja ' - . - ' '-v-' =(Tyvlu)

dp,;

This equality plus Equation (27.51) gives 2

2

A(u)lIvll = A(v)lluli

=}

A(V) A(U) 11vII2 = lIull 2'

Since lu) and Iv) are arbitrary, we conclude that A(U) = clluf for all la) E J/:, where c is a constant. Equations (27.51) and (27.52) now yield Equation (27.50). If we let lu) = Iv) in Equation (27.52) and use (27.51), we obtain

fa I(ul r, lu)1 2 a,.x = A(u)lIuIl

2

= cllull 4 .

That c > 0 follows from the fact that the LHS is a nonnegative continuous function that has at least one strictly positive value in its integration range, namely at x = e, the identity. D 27.3.9. Theorem. Every irreducible unitary representation of a compact Lie group is finite-dimensional. Proof Letllei) Ji'=l be any set of orthonormal vectors in J{. Then, unitarity ofT. implies that {T. lei)Ji'=1 is also an orthonormal set. Applying Lemma 27.3.8 to lej) and lei). we obtain

fa l(ellTx lej)12 d,.x = c. Now sum over j to get nc

= ~ fa I(ell r, lej)12 d,.x = fa ~ l(e1l Txej)1 2 d,.x

.s fa (ellel) d,.x

= VG,

where we used the Parseval inequality [Equation (5.3)] as applied to the vector leI) and the orthonormal set {T. lei)}i'=l' Since both VG and c are finite, n must be finite as well. Thus, J/: cannot have an infinite set of orthonormal vectors. D

852

27. LIE GROUPS ANO LIE ALGEBRAS

So far. we have discussed irreducihle representations. What can we say about arbitrary representations? We recall that in !he case of finite groups. every repre-

sentation can be written as a directsumof irreducible representations. Is thisalso true for compact Lie groups? Firstly. we note !hat !he Weyl operator. being Hilbert-Schmidt. is necessarily compact. It is also hermitian. Therefore. by the spectral theorem, its eij&enspaces span !he carrier space JL Specifically. we can write X = Mo Ell Lj=l EIl:M j. where Mo is !he eigenspace corresponding to !he zero eigenvalue of Ku• and N could be infinity. Secondly. from !he relation (vi Ku [u) = cllull 211vll 2 and !he fact !hat c "i- 0 and lu) "i- O.we conclude !hat Ku cannot have any nonzero eigenvector for its zero eigenvalne. It follows !hat Mo contains only !he zero vector. Therefore. if X is infinite-dimensional. !hen N = 00. Thirdly. consider any representation T of G. Because Ku commutes wi!h all Tg. each eigenspace of Ku is an invariant subspace under T. If a subspace 'U is invariant under T. !hen 'U n:Mi» a subspace of:M j. is also invatiant (reader. please verify!). Thus. all invatiant subspaces of G are reducible to invariant subspaces of eigenspaces of Ku. In particular. all irreducible invatiant subspaces of T are subspaces of eigenspaces of Ku. Lastly. since all :Mj are finite-dimensional. we can use !he procedure used in !he case of finite groups and decompose :Mj into irreducible invatiant subspaces of T. We have just shown !he following result: 27.3.10. Theorem. Every unitary representation T of a compact Lie group G is a direct sum of irreducible finite-dimensional unitary representations. By choosing a basis for !he finite-dimensional invatiant subspaces of T. we can represent each Tg by a matrix. Therefore.

27.3.11. Box. Compact Lie groups can be represented by matrices.

As in !he case of finite groups. one can work wi!h matrix elements and characters of representations. The ouly difference is !hat summations are replaced wi!h integration and order of the group IGI is replaced wi!h Va. which we take to be unity.? For example. Equation (24.6) becomes

fa T(a)(g)XT(~)(g-l)

d/Lg =

Ax8a~1.

(27.53)

and !he analogue of Equation (24.8) is

t

a

",(a) ( )

'it

(~)*

1

g Tj m (g) d/Lg = -8mI8a~8ij.

9Thiscanalwaysbe doneby rescaling thevolumeelement.

na

(27.54)

27.3 REPRESENTATION OF COMPACT LIE GROUPS

853

Characters satisfy similar relations: Equation (24.11) becomes

fa

X (a)(g)X (P)* (g) dJ.Lg = 8ap,

(27.55)

and the useful Equation (24.16) turns into

l

IX (g) ,2 dJ.Lg =

G

L>~'

(27.56)

a

This formula can be used to test for irreducibility of a representation:If the integral is unity, the representation is irreducible; otherwise, it is reducible. Finally,we state the celebrated Peter-Weyl theorem (for a proof, see [Barn 86, pp.I72-173]) Peter-Weyllheorem

27.3.12. Theorem. (Peter-Weyl theorem) The functions

s t. j s n a, form a complete set offunctions in .c 2 (G), the Hilbert space of square-integrable ,In; 1';;a) (g),

Va and I

functions on G.

If u E

.c2 (G ), we can write where

b5 = n a

1

u(g)1';;a)*(g) d/Lg.

G

(27.57)

27.3.13. Example. Equation (27.57) isthegeneralization oftheFourier series expansion

The Peter-Weyl theorem isthe generelization ofthe Fourier series

expansion ofperiodic functions.

of functions. The connection with Fourier series becomes more transparent if we consider a particular compact.group. The unit circle 8 1 is a one-dimensional abelian compact 1parameter Lie group. In fact, fixingan "origin" on the circle, any other point can be described by the parameter 8. the angular distance from the point to the origin. 8 1 is obviously abelian; it is also compact, because it is a bounded closed region of IR.2 (BWHB theorem). By Theorem 24.2.3, which holds for all Lie groups, all irreducible representations of 8 1 are I-dimensional. So T(~)(g) -+ T(a)(e). Furthermore, T(a)(e)T(a)(e') = T(a)(e + e'). IJ

Differentiating both sides with respect to 8' at 0' = 0 yields the differential equation

T

(a)

dT(a)

(e) - , -

I

de

0'=0

dr(a)

= --

dy

I

_ dT(a)

=--,

de

y=O

y=e+e'.

The solution to this DE is AeaB. Since T(a) are unitary, and since a I-dimensional unitary matrix must look like i fIJ • we must have A 1.Furthermore, and ()+ 21fare identified on the unit circle; therefore, we must conclude that a is i times an integer n, which determines the irreducible representation. We label the irreducible representation by n and write

=

T(n) (e) = einO,

e

n = 0 ± 1 ± 2 ....

ThePeter-Weyl theoremnowbecomesthe ruleof Fourierseriesexpansionof periodicfunctions. This last property follows from the fact that any function u : SI -+ JR is necessarily II periodic.

854

27. LIE GROUPS ANO LIE ALGEBRAS

There are many occasions in physics where the state functions describing physical quantities transform irreducibly under the action of a Lie group (which we assume to be compact). Often this Lie group also acts on the underlying spacetime manifold. So we have a situation in which a Lie group G acts on a Euclidean space R n as well as on the space of(square-integrable) functions L
O'(x', y', z') = O(x, y, z), where (x', y', z') describes the same point for 0' that (x, y, z) describes for O. In the context of representation theory, we can think of (x', y', z') as the transformed coordinates obtained as a result of the action of some group: (x', s'. z') = g . (x, y, z), or x' = g . x. So, the equation above can be written as O'(x') = O(x) = O(g-t. x)

or

O'(x) =O(g-l.x).

It is natural to call 0' the transform of 0 under the action of g and write 0' = TgO. This is one way of constructing a represeutatiou [see the comments after Equation (24.1)]. Instead of g-l on the left, one could act with g on the right. When the physical quantity is not a scalar, it is natural to group together the smallest set of functions that transform into one another. This leads to the set of functions that transform according to a row of an irreducible representation of the group. In some sense, this situation is a combination of (24.1) and (24.37). The reader may verify that n.

Tg¢;(a)(x) = LTj~a)(g)¢;a)(x.g-l)

(27.58)

j=l

defines a representation of G. We now use Box 27.1.26 to construct an irreducible representation of the Lie algebra of G from Equation (27.58). By the definition of the infinitesimal action,

855

27.3 REPRESENTATION OF COMPACT LIE GROUPS

we let g yields

fr

= exp(et) and differentiate both sides with respect to t at t = O. This

a) TexP(Et):i (x)

1,=0, = ~ :t (TJa\exp(et))¢t)(x . exp( -em} 1,=0

=Lj ~ji(~,)¢~OI)(X)

where we have defined the matrices :l) ji (e) for the LHS. The derivative in the first sum is simply '.rJ~) (e) the representation of the generator e of the l-parameter a

gronp of transformations in the space of functions {¢i ) }. The derivative in the second sum can be found by writing x' (t) = x . exp( -et) and differentiating as follows:

where we used Equation (27.18) and defined Xk(x; e) by the last equality. We also changed the coordinate index to Greek to avoid confusing it with the index of the functions. Collecting everything together, we obtain

(27.59) where X" (x; e) is the vth component of the infinitesimal generator of the actiou induced by e E g. We shall put Equation (27.59) to good use when we discuss symmetries and conservation laws in Chapter 30. The derivative with respect to the functions, although meaningless at this point, will be necessary when we discuss conservation laws.

856

27. LIE GROUPS AND LIE ALGEBRAS

27.4 Representation of the General Linear Group G L (V) is not a compact group, but we cau use the experience we gained in the aualysis of the symmetric group to find the irreducible representations of G L (V). The key is to construct tensor product spaces of V-which, as the reader may verify, is a carrier space of G L (V)-aud look for its irreducible subspaces. In fact, if r is au arbitrary positive integer, T : G --* G L (V) is a representation, aud V®, ==V®· .. ®V. ~ T

times

Then T®' : G --* GL(v®r), given by [T®r(g)](Vt, ... , vr ) sa T~r (VI, ... , vr ) = Tg(VI) ® ... ® Tg(v,), is also a representation. In particular, considering V as the (natural) carrier space for GL(V), we conclude that T®r : GL(V) --* GL(v®r) is a representation. This tensor product representation is reducible, because as is evident from its definition, T~' preserves auy symmetry of the tensor it acts on. For example, the subspace of the full n r -dimensional tensor product space consisting of the completely symmetric tensors of the type

t, sa

L

VnCI) ® Vn (2) ® ... ® vnC')

1rESr

is invariaut. Similarly, the subspace consisting of the completely autisymmetric tensor products-the r-fold wedge products-is invariant To reduce V®', we choose a basis {eklk=1 forV. Then the collection of nr tensor products {ek, ® ... ®ek,}, where each k i runs from 1 to n, is a basis for V®'. An invariaut subspace of v®r is a spau of linear combinations of certain of these basis vectors. Since the only thing that distinguishes among {ek, ® ... ® ek,} is a perroutation of the r labels, we star! to see the connection between the reduction of v®r aud S,. This conoection becomes more evident if we recall that the left multiplication of the group algebra of S; by its elements provides the regular representation, which is reducible. The irreducible representations are the minimal ideals of the algebra generated by the Young operators. The same idea works here as well: Certain linear combination of the basis vectors of v®r obtained by perroutations cau serve as the basis vectors for irreducible representations of GL(V). Let us elaborate on this. Recall that a Young operator of S; is written in the forro Y = QP where Q aud P are linear combinations of perroutations in Sr. Y has the property that if one operates on it (via left multiplication) with all perroutations of S" one generates a minimal ideal, i.e., au irreducible representation of S,. Now let Y be a Young operator that acts on the indices (kl' ... , k,), giving linear combinations of the basis vectors of V®'. From the minimality of the ideal generated by Y aud the fact that operators in G L (V)

27.4 REPRESENTATION OF THE GENERAL LINEAR GROUP

857

permute the factors in ekj 181· . ·l8Iek, in all possible ways, it should now be clear that if we choose any single basis vectorek j l8l " •181 ek, , then Y(ekj l8l " 'l8Iek,) generates an irreducible representation of G L (V). We therefore have the following: connection between the Young tableaux and irreducible representations of GL(V)

27.4.1. Theorem. Let {eklk~1 be any basis for V. Let Y = QP be the Young operator of S; that permutes (and takes linear combinations of) the basis vectors {ekj 181 ... 181 ek,.}. Then for any given such basis vector, the vectors {T~rY(ekl 181···181 ek,) I g

E

GL(V)}

span an irreducible subspace ofVer.

A basis of such an irreducible representation can be obtained by taking into account all the Young tableaux associated with the irreducible representation. But which of the symmetry types will be realized for given values of n and r? Clearly, the Young tableau should not contain more than n rows, because then one of the symbols will be repeated in a column, and the Young operator will vanish due to the antisymmetry in its column indices. We can therefore restrict the partition (A) to (A) = 0,1, A2, ... , An),

Al + ... +An = r,

Let us consider an example for clarification. 27.4.2. Example. First,let n = r = 2. Thetensorproduct spacehas22 = 4 dimensions. Toreduce it, we consider the Young operators, whichcorrespond to e + (kl, kl) and e (kt. k2)· Let us denote these operators by VI and Vl. respectively. By applying each one to a generic basisvector ekl ® ek2' we can generate all the irreducible representations. The first operator gives VI (ekj

e ek,) =

ekj

e ek, + ek, ® ek!'

where kl and k2 can be 1 or 2. For kl = ka = 1, we get 2el ® ei. For kl = 1, k2 = 2, or kl = 2, k2 = 1, we get el ® e2 + ez ® ej , Finally, for kl = k2 = 2, we get 2e2 ® e2.

Altogether, we obtain 3 linearly independent vectors that arecompletely symmetric. When thesecond operator actson a generic basisvector, it gives Y2(ekl ® ek2) = ekl ® ek2 - ek2 ® ekl'

The only time that this is not zero is when kl and k2 aredifferent. hi either case. we get ±(el ® ez - C2 ® el)· This subspace is therefore one-dimensional. Thereduction of thetensor product spacecantherefore be written as

3D symmetric subspace

Next, let us consider the case of n = 2. r = 3. The tensor product space has 23 = 8 dimensions, Toreduce it, we needto consider all Young operators of 83. There arefour of these.corresponding to thefollowing tableaux:

858

27. LIE GROUPS ANO LIE ALGEBRAS

Let us denote these operators by Vb Y2, Y3. and Y4,respectively. By applying each one to a generic basis vector ekl ® ek2 ® ek3' we can generate all the irreducible representations.

The first operator gives VI (ekl ® ek2 ® ek3)

= ekl

® ek2 ® ek3 + ekl ® ek3 ® ekz

+ ek2 ® ekl

® ek3

+~®~®~+~®~®~+~®~®~,

where zj , kz, andkz can be 1 or 2. Forkl = kZ = k3 = I, we get Sej ®el ®el·Forthe case where two of the ki'S are 1 and the third is 2, we get

For the case where two of the ki'S are 2 and the third is I, we get

Finally, for kl = kz = k3 = 2, we get 6ez ® ez ® ez. Altogether, we obtain 4 linearly independent vectors that are completely symmetric. When the second operator acts on a generic basis vector, it gives lO Yz(ek, ® ek,

e ek,) =

[e - (kl, k3)][e

= ekl ® ek2 ® ek3

+ (kl, kz)] (ek, e ek, ® ek,) + ekz ® ekl ® ek3

-ek3 ®ekz ®ekl -ek2 ®ek3 ®ekt"

If all three indices are the same, we get zero. Suppose kl = 1. Then kz can be 1 or 2. For k2 = 1, we must set k3 = 2 to get e2 ® ei ® ei - ej ® e2 ® ej . For k2 2, we must set k3 = 1 to obtain el ® ez @ e2 - ez ® ei ® ez. If we start with kl = 2, we will not produce any new vectors, as the reader is urged to verify. Therefore, the dimension of the irreducible subspace spanned by the second Young tableau is 2. The action of the third operator on a generic basis vector yields

=

Y3(ek, e ek, ® ek,) = [e - (kl. kz)][e + (klo k3)](ek, ® ek, ® ek,) = ekt ® ek2 ® ek3

+ ek2 ® ekl

® ek3

-ek3 ® ek2 ® ekl - ek3 ® ekl ® ek2'

The reader may check that we obtain a two-dimensional irreducible representation spanned by ej ® ej ® ez - ez ® el ® el and "I ® eZ ® ez - eZ ® ez ® ej. The fourth Young operator gives zero because it is completely antisymmetric in three slots and we have only two indices. The reduction of the tensor product space can therefore

lOWhen a symmetric group is considered as an abstract group-as opposed to a group of transformations-e-we may multiply permutations (keep track of how each number is repeatedly transformed) from left to right. However, since the permutations here act on vectors on their right, it is more natural to calculate their products from right to left.

27.5 REPRESENTATION OF LIE ALGEBRAS

859

be written as

dim=4

dim=2

Ell Span{Y3(ek, <8l ek, <8l ek,)} . dim=2

Wenotethat thetotal dimensions on both sidesmatch.

III

There is a remarkable formula that gives the dimension of all irreducible representations of GL(V) (see [Harne 89, pp. 384-387] for a derivation): 27.4.3. Theorem. Let V be an n-dimensional vector space, and V(A), the irreducible subspace of tensors with symmetry associated with the partition (A) = (AI, ... , An). Then dim V(A) =

D(l[, ... , In)

,

D(n - 1, n - 2, ... , 0) where Ij sa Aj + n - j and D(x[, ... , x n) is as given in Equation (24.56).

27.5 Representation of Lie Algebras The diffeomorphism established by the exponential map (Theorem 27.1.18) reduces the local study of a Lie group to that of its Lie algebra.'! In this book, we are exclusively interested in the local properties of Lie groups, and we shall therefore confine ourselves to Lie algebras to stody the structure of Lie groups. Recall that any Lie group homomorpbism leads to a corresponding Lie algebra homomorphism [Equation (27.7)]. Conversely, a homomorphism of Lie algebras can, through the identification of the neighborhoods of their identities with their Lie algebras, be "exponentiated" to a (local) homomorphism of their Lie groups. This leads to the following theorem (see [Fult 91, pp. 108 and 119] for a proof). 27.5.1. Theorem. LetGbeaLiegroupwithalgebrag.ArepresentotionT: G--> GL(Jf) determines a Lie algebra representation T. : g --> gl(Jf). Conversely, a Lie algebra representation X : g --> gl(Jf) determines a Lie group representation. It follows from this theorem that all (local) Lie group representations result from corresponding Lie algebra representations. Therefore, we shall restrict ourselves to the representations of Lie algebras. 11 Weuse theword"local" to meanthecollectionof allpointsthat canbe connected to theidentity by a curve in theLie group G.lfthis collectionexhausts G. thenwe say thatG is connected. If, furthermore, all closed curves (loops) in G canbe shrunk to a point, we say that G is simply connected.The word "local" canbe replaced by "simply connected" in what follows.

860

27. LIE GROUPS ANO LIE ALGEBRAS

27.5.1 Representation of Subgroups of GL(V) Let 9 be any Lie algebra with basis vectors {X;}. Let a representation T map these vectors to {Ti} E gl(~ for some carrier space Je. Then, a general element X = Li aiXi of 9 will be mapped to T = Li aiTi. Now snppose that ~ is a snbalgebra of g. Then the restriction of T to ~ provides a representation of~. This restriction may be reducible. If it is, then there is an iuvariant subspace Jet of Je. It follows that

(blTx 10) = 0

vX

E

~

whenever 10)

E

Jet and Ib)

E

Jet,

where Tx '" T(X). If we write Tx = Li a;X)Ti, then iu terms ofTi, the equation above can be written as dUng

L

dUng

a;X) (bl r, 10) sa L

i=1

a;X\(ba) = 0

VXE

~,

(27.60)

;=1

where ,?a) sa (bl Ti 10) are complex nmnbers. Equation (27.60) states that

27.5.2. Box. If T, as a representation of ~ (a Lie subalgebra of g), is re-

ducible, then there exist a number ofequations that a;X) must satisfy whenever X. E ~. If T, as a representation of g, is irreducible, then no relation such as given in (27.60) will exist when X runs over all of g.

This last statement will be used to analyze certain subgroups of GL(V). Let us first identify GL(V) with GL(n, iC). Next, consider GL(n, R), which is a subgroup of G L(n, iC), and transfer the discussion to their respective algebras. If {Xi} is a basis of g[(n, iC), then an arbitrary element can be written as Li aiXi. The difference between g[(n, iC) and gl(n, R) is that the ai's are real iu the latter case; i.e., for all real values of {ai}, the sum belongs to g[(n, R). Now suppose that T is an irreducible representation of g[(n, iC) that is reducible when restricted to gI(n, R). Equation (27.60) states that the function n2

f(Zt, ... , zn2) sa LZi,,lba) ;=1

vanishes for all real values of the Zi'S. Siuce this function is obviously entire, it must vanish for all complex values of Zi'S by analytic continuation (see Theorem 11.3.1). But this is impossible because T is irreducible for g[(n, iC). We have to conclude that T is irreducible as a representation of g[(n, R). The next subalgebra of gI(n, iC) we consider is the Lie algebra .[(n, iC) of the specialliuear group. The only restriction on the elements of .[(n, C) is for them

27.5 REPRESENTATION OF LIE ALGEBRAS

861

to have a vanishing trace. Denoting tr X, by ti, we conclude that X = Li O!iXi belongs toal(n, C) if and only ifLi O!iti = O. Let (tt, ... , t:,) es It} E en'. Then

a[(n, C) can be characterized as the subspace consisting of vectors la} E en' such that (al t) = O. Such a subspace has n 2 - I dimensions. If any irreducible representationofgl(n, C) is reducible foral(n, C), then the set ofcomplex numbers 100i} must, in addition, satisfy Equation (27.60). This amounts to the condition that la} be orthogonal to Ir(bal} as well. But this is impossible, because then the set lla} , It} , Ir(bal}} would constitute a subspace of en' whose dimension is at least n 2 + I: There are n2 - I of la),s, one It}, and at least one Ir(bal}. Therefore, all irreducible representations of glen, C) are also irreducible representations of a[(n, iC). The last subalgebra of glen, C) we consider is the Lie algebra u(n) of the unitary group. To study this algebra, we star! with the Weyl basis of Equation (27.27) for glen, C), and construct a new hermitian basis IXkj} defined as Xjj ==ejj Xkj

== i(ekj - ei)

forallj = 1,2, ... ,n,

if k

01 j.

A typical element of glen, C) is of the form Lkj O!kjXkj, where O!kj are complex numbers. If we restrict ourselves to real values of O!kj, then we obtain the subalgebra of hermitian matrices whose Lie group is the unitary group U(n). The fact that the irreducible representations of glen, iC) will not reduce under u(n) follows immediately from our discussion concerning glen, R). We summarize our findings in the following:

27.5.3. Theorem. The irreducible representations of GL(n, C) are also irreducible representations ofGL(n, R), SL(n, ci U(n), and SU(n). The case of SU (n) follows from the same argument given earlier that connected GL(n, iC) to SL(n, iC).

27.5.2 CasimirOperators In the general representation theory of Lie algebras, it is desirable to label each irreducible representation with a quantity made out of the basis vectors of the Lie algebra, An example is the labeling of the energy states of a quantum mechanical system with angular momentum. Each value of the total angular momentum labels an irreducible subspace whose vectors are further labeled by the third component of angular momentum (see Chapter 12). This subsection is devoted to the generalization of this concept to an arbitrary Lie algebra. Casimir operator defined

27.5.4. Definition. Let X: g -+ g[(1f) be a representation of the Lie algebra g. A Casimir operator C for this representation is an operator that commutes with all TX of the representation. If the representation is irreducible, then by Schur's lemma, C is a multiple of the unit operator. Therefore, all vectors of an irreducible invariant subspace of

862

27. LIE GROUPS AND LIE ALGEBRAS

Chevalley's theorem

the carrier space J{ are eigenvectors of C corresponding to the same eigenvalue. That Casimir operators actually determine the irreducible representations of a semisimple Lie algebra is the content of the following theorem (for a proof, see [Vara 84, pp. 333-337]). 27.5.5. Theorem. (Chevalley) For every semisimple Lie algebra lJofrank12 r with a basis {Xi}, there exists a set of r Casimir operators in the form ofpolynomials in TXi whose eigenvalues characterize the irreducible representations of g.

From now on, we shall use the notation Xi for TXi' It follows from Theorem 27.5.5 that all irreducible invariant vector subspaces of the carrier space can be labeled by the eigenvalues of the r Casimir operators. This means that each invariant irreducible subspace has a basis all of whose vectors carry a set of r labels corresponding to the eigenvalues of the r Casimir operators. One Casimir operator-in the form of a polynomial of degree two-which works only for semisimple Lie algebras, is obtained easily: C = LgijXiXj,

(27.61)

i,j

where gij is the inverse of the Cartan metric tensor. In fact, with the summation convention in place, we have [C, Xkl = gij[XiXj, Xkl = gij {Xi [Xj, Xk] + [Xi, Xk]Xj}

+ erkXrXj} gij crk (XjXr + XrXj) gijg" Oiks (XjX, + X,Xj)

= gij Icjk XiXr

= =

= O.

(because gij is symmetric)

(because gij gsr Ciks is antisymmetric in j, r)

The last equality follows from the fact that gij and g" are symmetric, Ciks is completely antisymmetric [see the discussion following Equation (27.45)], and there is a som over the dummy index s. 27.5.6. Example. Therotation group SO(3)

in]R3

is a compact 3-parameter Lie group.

The infinitesimal generators arethe three components of the angular momentum operator (seeExample 27.1.29).From thecommutation relations of theangularmomentumoperators developed in Chapter 12, we conclude that c~ = iEijk. It follows thatthe Caftan metric tensor is gij =

ciscjr =

(iEisr )(iEjrs)

= +EisrE jsr = 2liij.

Ignoring the factorof 2 anddenotingthe angular momentum operators by Lj,we conclude

that

L2 == LI +L~ +L~ 12Recall that therank of fI is the dimension of theCartan subalgebra of fl.

27.5 REPRESENTATION OF LIE ALGEBRAS

irreducible representations of the rotation group and spherical harmonics

Connection between Casimir operators

and the PDEs of mathematical physics

863

is a Casimir operator. But this is precisely the operator discussed in detail in Chapter 12. Wefound there that theeigenvalues of L2- werelabeled by j, where j waseither aninteger or a half odd integer. In the contextof ourpresent discussion, we note thattheLie algebra 50(3) hasrank one,becausethere is no higher dimensional subalgebra of so(3) all of whose vectors commute with one another. It follows from Theorem 27.5.5 that L2 is the only Casimir operator, and that all irreducible representations T(j) of so(3) are distinguished by their label j. Furthermore, the construction of Chapter 12 showed explicitly thatthe dimension of T(J) is 2j + 1. Theconnection betweentherepresentation of Lie algebras andLie groups permits us to conclude thatthe irreducible representations of the rotation group are labeledby the (half) integers i. andthe jth irreducible representation hasdimension 2j + 1.Whenj is aninteger 1 andthecarrier spaceis /:.;2(S2), thesquare-integrable functions on theunitsphere, thenL2 becomes a differential operator, andthe spherical harmonics Ylm «(), rp), with a fixedvalue II of l, provide a basisfor the lth irreducible invariant subspace.

The last sentence of Example 27.5.6 is at the heart of the connection between symmetry, Lie groups, and the equations of mathematical physics. A symmetry operation of mathematical physics is expressed in terms of the action of a Lie group on an underlying manifold M, i.e., as a group of transformations of M. The Lie algebra of such a Lie group consists of the infinitesimal generators of the corresponding transformation. These generators can be expressed as first-order differential operators as in Equation (27.20). It is therefore natural to choose as the carrier space of a representation the Hilbert space £,2 (M) of the square-integrable functions on M, which, through the local identification of M with jRm (m = dim M), can be identified with functions on jRm • Then the infinitesimal generators act directly on the functions of £,2(M) as first-order differential operators. The Casimir operators (Ca}~~l' where r is the rank of the Lie algebra, are polynomials in the infinitesimal generators, i.e., differential operators of higher order. On the irreducible invariant subspaces of £,2(M), each Ca acts as a multiple of the identity, so if f(r) belongs to snch an invariant subspace, we have

Caf(r) = )..(OI)f(r),

ex. = 1,2, ... , r.

(27.62)

This is a set of differential equations that are invariant under the symmetry of the physical system, i.e., its solutions transform among themselves under the action of the group of symmetries. It is a stunning reality and a fact of profound significance that many of the differential equations of mathematical physics are, as in Equation (27.62), expressions of the invariance of the Casimir operators of some Lie algebra in a particular representation. Moreover, all the standard functions of mathematical physics, such as Bessel, hypergeometric, and confluent hypergeometric functions, are related to matrix elements in the representations of a few of the simplest Lie gronps (see [Mill68) for a thorough discussion of this topic).

Claude Chevalley (1909-t984) was the only son of Abel and Marguerite Chevalley who were the authors of the Oxford ConciseFrench Dictionary. He studied under Emile Picard

864

27. LIE GROUPS ANO LIE ALGEBRAS

at the Ecole Normale Superieur in Paris, graduating in 1929 and becoming the youngest of the mathematicians of the Bourbaki school. After graduation, Chevalley went to Germany to continue

his studies under Artin at Hamburg during the session 1931-32. He then went to the University of Marburg to work with Hasse, who had been appointed to fill Hensel's chair there in 1930. He was awarded his doctorate in 1937. A year later Chevalley went to the Institute for Advanced Study at Princeton, where he also served on the faculty of Princeton University. From July 1949 until June 1957 he served as professor of mathematics at Columbia University, afterwards returning to the University of

Paris. Chevalley had a major influence on the development of several areas of mathematics. His papers of 1936aod 1941led to major advaocesin class fieldtheory aod also in algebraic geometry.He did pioneering work in the theory of local rings in 1943, developingthe ideas of Krull into a theorem bearing his name. Chevalley's theorem was important in applications made in 1954to quasi-algebraically closed fieldsaod the followingyear to algebraicgroups. Chevalleygroupsplaya central role in the classificationof finite simple groups. His name is also attached to Chevalleydecompositions and to a Chevalley type of semisimple algebraic group. Many ofhis texts havebecome classics. He wrote Theory ofLie Groups in three volumes which appearedin 1946,1951, aod 1955.He also published Theory ofDistributions (1951), Introduction to the Theory ofAlgebraic Functions of one Variable (1951), The Algebraic Theory of Spinors (1954), Class Field Theory (1954), Fundamental Concepts ofAlgebra (1956), aod Foundations ofAlgebraic Geometry (1958). Chevalley was awarded many honors for his work. Among these was the Cole Prize of the AmericanMathematicalSociety.He was elected a member of the London Mathematical Society in 1967.

27.5.3 Representation of .50(3) and .50(3, 1) Because of their importance in physical applications, we study the representations Of50(3), the rotation, and50(3, I), the Lorentz, algebras. For rotations, we define Jl '" -iM23, J2 '" iMB, and J3 '" -iM12Y and note that the J, 's satisfy exactly the same commutation relations as the angular momentum operators of Chapter 12. Therefore, the irreducible representations of 50(3) are labeled by j, which can be an integer or a balf-odd integer (see also Example 27.5.6). These representations are finite-dimensional because 80(3) is a compact gronp (Example 27.3.3 and Theorem 27.3.9). The dimension of the irreducible representation of50(3) labeled by j is 2j + 1. Because of local isomorphism of Lie groups and their Lie algebras, the same irreducible spaces found for Lie algebras can be used to represent the Lie groups. 13Sometimes we use J x• J y, and J z instead of Jl, J2, and J3.

27.5 REPRESENTATION OF LIE ALGEBRAS

865

In particular, the states (Ijm) }~,=_j' where m is the eigenvalue of J" can also be used as a basis of the j -th irreducible representation. The flow of each infinitesimal generator of 50(3) is a one-parameter subgroup of S 0(3). For example, exp(M12rp)is a rotation of angle 'I' about the z-axis. Using Euler angles, we can write a general rotation as R(1f!,

e, '1') =

exp(M121f!) exp(M31e) exp(M12rp).

The corresponding rotation operator acting on the vectors of the carrier space is R( 1f!, e, '1') = exp(M121f!) exp(M31 e) exp(M12rp) rotation matrix

= eiJ,Y,eiJ,e eiJ,~.

The rotation matrix corresponding to the above operator is obtained by sandwiching R(1f!, e, '1') between basis vectors of a given irreducible representation:

D~,~(1f!, e, '1') '" (jm'l R(1f!, e, '1') Ijm) = (jm'l eiJ,Y, eiJye eiJ,~ Ijm) = eim'Y, eim~ (jm'l eiJ,e Ijm) = ei(m'y,+m.)d;;;'~(e).

(27.63)

Wigner formula for rotafion matrices

Thus, the calculation of rotation matrices is reduced to finding d;;;'~ (e). These are given by the Wigner formula (see [Hame 89, pp. 348-357]):

.

( e)2(j-JL)+m-m

d~,~(e) = Lq,(j,m,m';!-') cos2" "

1

e)2JL +m - m l

(

sin2

(27.64)

where

. m m': = -1" [(j + m)!(j - m)!(j + m')!(j - m')!]1/2 q,(j, , ,!-,) - ( ) (.j + m -!-' )1.!-'.1('j - m I -!-')'( . m I - m +)1 !-'. and the summation extends over all integral values of !-' for which the factorials have a meaning. The number of terms in the summation is equal to 1 + r, where r is the smallest of the four integers j ± m, j ± mi. From the rotation matrices, we can obtain the characters of the rotation group. However, an easier way is to use Euler's theorem (Theorem 4.7.7), Example 23.2.19, and Box 24.2.7 to conclude that the character of a rotation depends only on the angle of rotation, and not on the direction of the rotation axis. Choosing the z-axis as ouronly axis of rotation, we obtain X(j)(rp) =

j

j

L

(jml eiJ,~ Ijm) = L

m=-j

..

j

eim~ = e- iN L ei(j+m)~ m=- j m=-i

2j.

..

ei(2j+l)~

_ 1

= e-lJrp '"'" e lkrp = e-1j'P"---,-;:.,----,--L.., e1f/J - 1

=

k=O ei(j+l)~ _ e-ij~ ei~

- 1

=

sin(j

+ ~)

sin(rpj2)

(27.65)

866

27. LIE GROUPS ANO LIE ALGEBRAS

Equation (27.65) can be used to obtain the celebrated addition theorem for angular momenta. Suppose that initially we bave two physical systems corresponding to angular momenta l: and h- When these systems are made to interact with one another, the total system will be described by the tensor product states. These states are vectors in the tensor product of the irreducible representations T Ch) and TCh) of the rotation group, as discussed in Section 24.7. This product is reducible. To find the factors into which it reduces, we consider its character corresponding to angle tp, Using Equation (24.44), we have XUtxh)(rp)

it

L

= X Cill (rp) . X ( 2)(rp) =

eimt~

mt=-it h:

h

L L

=

iI+h

L

=

eimz~

m2=-h

e i (m I +m2)rp

J

L

eiM~ =

1=lh-hIM=-1

addition theorem for angular momenta

12

L

h+h

L

XCI)(rp),

1~lh-hl

where the double sum on the third line is an equivalent way of writing the double summation of the second line, as the reader may verify. From this equation we read off the Clebsch-Gordan decomposition of the tensor product: TUt>

e TUz) =

h+h

L

TCI),

(27.66)

1=lh-hl

which is also known as the addition theorem for angular momenta. Equation (27.66) shows that (see page 702)

27.5.7. Box. The rotation group is simply reducible.

The RHS of Equation (27.66) tells us which irreducible representations result from multiplying TUt) and TUZ). In particular, if h = iz == I, the RHS includes the J = 0 representation, i.e., a scalar. In terms of the states, this says that we can combine two states with angular momentum I to obtain a scalar state. Let us find this combination. We use Equation (24.48) in the form

L

IJM) =

C(hh Jlmj m2; M) Ih, mj; h, m2),

mj +m2 =M.

ml,m2

In the case under investigation, J = 0 = M, so (27.67) becomes I

100) =

L m=-l

C(ll; Dim, -m; 0) 11m; I, -m) .

(27.67)

867

27.5 REPRESENTATION· OF LIE ALGEBRAS

Problem 27.33 shows that C(ll; Dim, -m; 0) = (_I)I-m / ,J21 + I, so that I

100) = .

(_l)l-m

L ,J2TTI 11m; I, -m) . m~-l 21+1

Take the "inner product" of this with (IJ, tp; IJ', ",'I to obtaio

(IJ, tp; IJ', ",'I 00) =

=

I

(_l)l-m

LI

(_l)l-m

L ,J2TTI (IJ, tp; IJ', ",'11m; I, -m) m~-l 21 + I I

m=-

,

,

,J2TTI(IJ,,,,llm)(IJ,,,,II,-m), 21 + I ' " - - - v o - ' ' - v - ' Ylm(e,~)

(27.68)

Yl,-m«()',ql)

where we have used (IJ, ip; IJ', ",'I = (IJ, ",I (IJ', ",'I and contracted each bra with a ket. We can evaluate the LHS of (27.68) by noting that since it is a scalar, the choice of orientation of coordinates is immaterial. So, let IJ 0 to get IJ' y, the angle between the two directions. Then using the facts

=

Ylm(O, ",) = SmO

r:

and

--

4".

YlO(lJ, ",) =

=

Vfii+l ~PI(coslJ)

on the RHS of (27.68), we obtaio

(IJ, tp; IJ',

",'I 00)

(-Ii

= --,J21 4".

+ I PI(COSY).

Substituting this in the LHS of Equation (27.68), we get I

41l '"'" m ., ! PI (COS y) = 21 + I m~} -I) Ylm(lJ, "')YI,-m(1J , '" ),

which is the addition theorem for spherical harmonics discussed in Chapter 12. Let us now turn to .0(3, 1). We collect the generators in two categories M sa (Ml, M2, M3) sa (M23, M31, M12), N es (Nl, N2, N3) es (MOl, M02, M03),

and verify that [Mi, Mj] = -€ijkMk,

[NI, Nj] = €ijkMk,

[MI, Nj] = -€ijkNk,

and that there are two Casimir operators: M 2 - N 2 and M . N. It follows that the irreducible representations of .0(3, 1) are labeled by two numbers. To find these

numbers, define the generators

J ==

1 2i (M + iN),

1

.

K== -(M-IN) 2i '

8&8

27. LIE GROUPS AND LIE ALGEBRAS

and show that

It follows that the J's and the K's generate two completely independent Lie algebras isomorphic to the angular momentum algebras and that so(3, I) is a direct sum of these algebras. Since each one requires a (half-odd) integer to designate its

irreducible representations, we can choose these two numbers as the eigenvalues of the Casimir operators needed to label the irreducible representations ofso (3, I). Thus, the irreducible representations of so(3, I) are of the form TW'), where j and i' can each be an integer or a half-odd integer.

27.5.4 Representation of the Poincare Algebra The Poincare algebra p(p, n - p), introduced in Section 27.2.1, is the generalization of the Lie algebra of the invariance group of the special theory of relativity. It contains the Lorentz, the rotation, and the translation groups as its proper subgroups. Its irreducible representations are of direct physical significance, and we shall study them here. As the first step in the construction ofrepresentations of p(p, n - p), we shall try to find its Casimir operators. Equation (27.61) suggests one, but it works only for semisimple Lie algebras, and the Poincare algebra is not semisimple. Nevertheless, let us try to find an operator based on that construction. From the commutation relations for P(p, n - p), as given in Equation (27.40), and the double-indexed structure constants defined by,14

we obtain mn

,m,n

,m,n

Cjj,kl = OJ 01 tlik - OJ 0kTJa

+ Vi,m,n ukTJjl -

CV,k = 8j11ik - 8'('1}jk.

,moo

Uj

01 Tljt,

(27.69)

From these structure constants, we can construct a doubleindexed"metric" .. _ rs ...mn glj,kl - Cij,mn(;kl,rs

+ Cij,mCkl,T' r m

which the reader may verify to be equal to

14Pleasemakesureto differentiate between thepair(Mij, Pk) (which acts on p) andthe pair (Mij, Pk), which acts on the state vectors in theHilbert spaceof representation.

27.5 REPRESENTATION OF LIE ALGEBRAS

869

There is no natnral way of constructing a single-indexed metric. Therefore, we can only contract the M's. In doing so, it is understood that the indices are raised and lowered by ryij' So, the first candidate for a Casimir operator is ""kl ""kl"" M2 sa gij.klM'JM = 2(n - l)(ryjkryil - ryikryjl)M'JM = -4(n - I)M'JMij

The reader may verify that M2 commntes with all the Mij 's but not with the pi,s. This is to be expected because M2, the total "angular momentnm" operator'P is a scalar and should commute with all its components. But commutation with the pi'8 is not guaranteed. The construction above, although a failure, gives us a clue for a successful construction. We can make another scalar out of the P's. The reader may check that p 2 es ryijpiPj indeed commutes with all elements of the Poincare algebra. We have thus found one Casimir operator. Can we find more? We have exhausted the polynomials of degree two. The only third-degree polynomials that we can construct are Mij Pi Pj and ryilMijM jkMkl. The first one is identically zero (why?), and the second one will not commute with the P's. To find higher-order polynomials in the infinitesimal geuerators, we build new tensors out of them and contract these tensors with one another. For example,

consider the vector (27.70) Then c' Ci, a fourth-degree polynomial in the generators, is a scalar, and therefore, it commutes with the Mij 's, but unfortunately, not with Pi'S. Another common way to construct tensors is to contract various numbers of

the generators with the Levi-Civita teusor. For example, (27.71) is a contravariant tensor of rank n - 3. Let us contract W with itself to find a scalar (which we expect to commute with all the Mij's): W2 = Wit ...in- 3 W· -

.

lJ··· 1n- 3

-

Ei l ...i1l-3jkI M· P E" M"pt [k: I 1I···ln_3rst

,i2 = ( - l)n_" L..J Ejfu,it JT(it)u n(i2)

8i.- 3 8 j 8k 81 M P M"P' ... 1l"(in-3) nCr) n(s) net) jk I

n

= (-I)P(n - 3)!

L E,,8£(r)8~(,)8;(t)MjkPIM"pt, -n

where we used Equation (25.21). The sum above can be carried out, with the final result W2

= 2(-I)P(n - 3)! (MijM ijp2 - 2CiCi) = 2(-I)P(n - 3)! (M 2p2 - 2C2 ) ,

15This "angular momentum" includesordinary rotations as well as theLorentz boosts.

(27.72)

870

27. LIE GROUPS AND LIE ALGEBRAS

where Ci was defined in Equation (27.70). We have already seen that M2 , p 2 , and C2 all commute with the Mjk 's. The reader may check that W 2 commutes with the Pj 's as well. In fact,Wi t .. .in- 3 itself commutes with all the Pj 'so Other tensors and Casimir operators canbe constructed in a similar fashion. We now want to construct the irreducible vector spaces that are labeled by the eigenvalues of the Casimir operators. We take advantage of the fact that the poincare algebra has a commutative subalgebra,the translation generators. Since the Pk'S commute among themselves and with p 2 and W2, we can choose simultaneous eigenvectors of {PklZ~I' p2 , and W 2. In particular, we can label the vectors of an irreducible invariant subspace by the eigenvalues of these operators. The p 2 and w2 labels will be the same for all vectors in each irreducible invariant subspace, while the Pk'S will label different vectors of the same invariant subspace. Let us concentrate on the momentum labels and lei I1frt} be a vector in an irreducible representation of p(p, n - p), where p labels momenta and I-' distinguishes among all different vectors that have the same momentum label. We thus have

for k = 1,2, .... n,

(27.73)

where Pk is the eigenvalue of Pk. We also need to know how the "rotation" operators act on l1frt). Instead of the full operator eMijO'j, we apply its small-angle approximation 1 + Mijeij. Since all states are labeled by momentum, we expect the rotated state to have a new momentum label, i.e., to be an eigenstate of Pk. We want to show that (1 + Mije ij) l1frt} is an eignevector of Pk. Let the eigenvalue be p', which should be slightly different from p. Then, the problem reduces to determining 8p' es p' - p. Ignoring the index I-' for a moment, we have Pk

l1frp'}

= PI. l1frp') = (Pk + 8Pk) (1 + Mijeij)

l1frp) .

Using the commutation relations between Pk and Mij, we can write the LHS as LHS = Pk (1

+ Mijeij) l1frp} =

[Pk +eij (MijPk

+ ~jkPi -

~ikPj)] l1frp).

The RHS, to first order in infinitesimal quantities, can be expressed as RHS = (Pk

+ 8Pk + Pkeij Mij) l1frp} .

Comparison of the last two equations shows that ..

8Pk = e'} (~jkPi - ~ikPj)

..

= e'}

I"

(~jk~il - ~ik~jl) P

I

= e'} (Mij)kl P ,

where we used Equation (27.37). It follows that p' = p

+ 8p =

(1 + eijMij)p,

stating that the rotation operator of the carrier Hilbert space rotates the momentum label of the state. Note that since "rotations" do not change the length (induced by 'T/), p' and p have the same length.



27.5 REPRESENTATION OF LIE ALGEBRAS

construction and properties oftheIiltle group

871

To obtain all the vectors of an irreducible representation of p (p, n - p), we must apply the rotation operators to vectors snch as Itt). But not all rotations will change the label p; for example, in three dimensions, the vector p will not be affected16by a rotation about p. Tltis motivates the following definition.

27.5.8. Definition. Let Po be any given eigenvalue ofthe translation generators. The set 9
group 0 (p, n - p), called the little group corresponding to Po. The little algebra consists of the generators M~o satisfying - 0. MPO ijPo-

induced representations

The significance of the little group resides in the fact that a representation of 9
Tltis shows that (AP'PO) -I AAPPo belongs to the little group. So,

= AP'poAPO (APPtJ)-1 Itt) = AP'poAPO Itto) = AP'PtJ L: TV/L(APO) It;o) = L: TV/L(APO)AP'PO It;:')

T(A) Itt) sa A Itt)

v

=

v

L:v TV/L(APO) It;,) = L:v TV/L(APO) ItX p)'

Note how the matrix elements of the representation of the little group alone have entered in the last line. We therefore consider

16Thereader should be warnedthat although such a rotation does not change p. the rotation operator may change the state I'iflpf' However, the resulting statewill be aneigenstate of the Pk's with eigenvalue p. Weareusing the fact that 0 (p, n - p) is transitive (see Problem27.39).

872

27. LIE GROUPS AND LIE ALGEBRAS

To avoid confusion, we have used R for the representation of the little group. We claim that Equation (27.74) defines a (matrix) representation of the whole group. In fact,

L RVl'(A~) IlfrX2P} = L Rvl'(~O) L Rpv(A\",) Ilfrf,A2P} v

T(Aj)T(Az) Ilfr~} = T(Aj)

v

p

=

~ (~Rpv(A~O)RVl'(~)).llfrf'A2P)' =Rpp;(Aro~Po) since R is a rep.

ThereadermaycheckthatA~OA~O es (AjAz)Po. Therefore,

T(Aj)T(Az) Ilfr~} =

L Rpl' ((AjAz)PO) Ilfrf,A2P}

es T(AjAz)

Ilfr~} ,

p

and T is indeed a representation. It turns out that if R is irreducible, then so is T. The discussion above shows that the irreducible representations of the Poincare group are entirely determined by those of the little group and Equation (27.73). The recipe for the construction of the irreducible representations ofp(p, n - p) is now clear:

27.5.9. Theorem. Choose any simultaneous eigenvector Po of the Pk'S. Find the little algebra !Rpo at Po by finding all Mij's satisfying MijPO = O. Find all irreducible representations of !Rpo' The same eigenvalues that label the irreducible representations of!RPlJ can be used, in addition to those ofP z and WZ, to label the irreducible representations ofp(p, n - pl. We are particularly interested in p(3, I), the synnnetry group of the special theory of relativity. In applying the formalism developed above, we need to make contact with the physical world. This always involves interpretations. Borrowing from the angolar momentnm theory, in which a physical system was given the attribote ofangolar momentum, the label of the irreducible representation ofthe rotation group, we attribote the labels of an irreducible representation of the Poincare group, i.e., the eigenvalues of the four translation generators and the two Casintir operators, to a physical system. Since the four translation generators are identified as the three components of momentnm and energy, and their specification implies

their constancy overtime, we have to come to the conclusion that

27.5.10. Box. An irreducible representation ofthe Poincare group specifies afree relativistic particle.

27.5 REPRESENTATION OF LIE ALGEBRAS

873

There may be some internal interactions between constituents of a (composite) particle, e.g. between quarks inside a proton, but as a whole, the composite will be interpreted as a single particle. To construct the little group, we have to specify a 4momentum Po. We shall consider two cases: In the first case, Po . Po i' 0, whereby the particle is deduced to be massive and we can cheosel'' Po = (0,0,0, m). In the second case, Po . Po = 0, in which case the particle is massless, and we can choose Po = (p, 0, 0, pl. We consider these two cases separately. The little group (really, the little Lie algebra) for Po = (0, 0, 0, m) is obtained by searching for those rotations that leave Po fixed. This is equivalent to searching , for Mij's that annihilate (0, 0, 0, m), namely, the solutions to

(MijPo)/ = (Mij)/r(Po)r = (Mij)lOm =

° =>

(Mij)1O = 0.

°

°

Since (Mij)1O = ~iO~j/-~jO~il, we conclude that (Mij)1O = if and onlyifi i' and j i' 0. Thus the little group is generated by (MZ3, M31, M12) which are the components of angular momentum. The reader may also verify directly that when the 4-momentum has only a time component, the Casimir operator W2 reduces essentially to the total angular momentum operator. Since we are dealing with a single particle, the total angnlar momentum can only be spin. Therefore, we have the following theorem. 27.5.11. Theorem. In the absence ofany interactions, a massive relativistic particle is specified by its mass m and its spin s, the former being any positive number, the latter taking on integer or half-odd-integer values.

The case of the massless particle can he handled in the same way. We seek those Mij's that annihilate (p, 0, 0, p), namely, the solutions to (MijPO)k = (Mij)kr(PO)' = (Mijlkop

+ (Mij)klP =

0.

The reader may check that

(MOIPO)k = ~lkP - ~OkP,

(M02PO)k = ~2kP,

(MQ3Polk = ~3kP,

(MZ3PO)k = 0,

(MI2PO)k = ~2kP,

(M13Polk = ~3kP·

Clearly, MZ3 is one of the generators of the little group. Subtracting the middle terms and the last terms of each line, we see that Moz - MI2 and M03 - M13 are the other two generators. These happen to be the components ofW. In fact, it is easily verified that W O = Wi = MZ3P,

W Z = 2p(M13 - M03),

W 3 = 2p(Moz - M12). (27.75)

Therefore, the little group is generated by all the components of W. Furthermore, W2 has zero eigenvalue for IVrpo) when Po = (p, 0, 0, pl. Since both Casimir 18We use units in which c = 1.

874

27. LIE GROUPS AND LIE ALGEBRAS

operators annihilate the stale l1frpo)' we need to come up with another way of labeling the states.

Eugene Paul Wigner (t902-1995) was the second of three children born to Hungarian Jewish parents in Budapest. His father operated a large leather tannery and hoped that his son would follow him in that vocation, but the younger Wigner soon discovered both a taste and an aptitude for mathematics and physics. Although Wigner tried hard to accommodate his father's wishes, he clearly heard his calling, and the world of

physics is fortunate that he did. Wigner began his education in what be said "may have been the finest high school in the world." He later studied

chemical engineering and returned to Budapest to apply that training in his father's tannery. He kept track of the seminalpapers during the early years of quantum theory and, when the lure ofphysics became too strong, returned to Berlin to work in a crystallography lab. He lectured briefly at the University of Gfittingen before moving to America to escape the Nazis. Wigner accepted a visiting professorship to Princeton in 1930. When the appointment was not made permanent, the disappointed young professor moved to the University of Wisconsin, where he served happily until his new wife died suddenly of cancer only a few months after their marriage. As Wigner prepared, quite understandably, to leave Wisconsin, Princeton corrected its earlier mistake and offered him a permanent position. Except for occasional visiting appointments in America and abroad, he remained at Princeton until his

death. Wigner's contributions to mathematical physics began during his studies in Berlin, where his supervisor suggested a problem dealing with the symmetry of atoms in a crystal. John von Neumann, a fellow Hungarian physicist, pointed out the relevance of papers by Frobeniusand Schur on representation theory. Wigner soon became enamored with the group theory inherent in the problem and began to apply that approach to quantum mechanical problems. Largely at the urging of Leo Szilard (another Hungarian physicist and Wigner's best friend), Wigner collected many of his results into the classic textbook Group Theory and Its Application to the Quantum Mechanics ofAtomic Spectra. The decades that followed were filled with important contributions to mathematical physics, with applications of group theory comprising a large share: angular momentum; nuclear physics and SU(4) or "supermultiplet" theory; parity; and studies of the Lorentz group and Wigner's classic definition of an elementary particle. Other work included early efforts in many-body theory and a paper on level spacings derived from the properties of Hermitian matrices that later proved useful to workers in quantum chaos. As with most famous figures, Wigner's personality became as well known as his professional accomplishments. His insistence on ''reasonable'' behavior, for instance, made him refuse to pay a relative's hospital bill until after the patient was released-it was obviously umeasonable to hold a sick person hostage. His gentleness is exemplified in an anecdote in which on getting into an argument about a tip with a New York City cab driver, Wigner loses his patience, stamps his foot, and says, "Oh, go to hell, ... please!"

27.5 REPRESENTATION OF LIE ALGEBRAS

875

Heheldothers' feelings in such highregard thatit wassaidto be impossible to follow Wigner through a door. He waslight-hearted andfun-loving, but also devoted tohis family and concerned about the future of the planet. This combination of exceptional skill and

laudable humanity eusures Wigner's place among the mosthighly regarded of hisfield. (Taken from E. Vogt, Phys. Today 48 (12) (1995) 40--44.)

Define the new quantities 1 Ho= -Wo

2p

andthe correspondingoperatorsactingon the carrier space.From Equation (27.75), it follows that [WI, W2] = 0, W2 = -4H+H_, and that

Denote the eigenstates of W2 and Ho by la, 13): W

2

Ia , f3 ) = a la, 13)

,

HO la, 13) =

13 la, 13) .

Then the reader may check that H± la, 13) has eigenvalues a and 13 ± 1. By applying H± repeatedly, we can generate all eigenvaluesof Ho and note that they are of the foun

13 = r + n,

n = 0, ±l, ±2, ...

where

and

1 > r 2: O.

Since Ho = M23, Ho is recognized as an angnlar momentum operator whose eigenvalues are integer (for bosons) and half-odd integer (for feunions). Therefore, r = 0 for bosons and r = for feunions. Now, within an irreducible representation, only those la, 13) 's can occur that have the sarae a. Therefore, if we relabel the 13 values by integers, then

!

(a, nl Ho la, m) = (r

+ n)8nm •

Similarly, (a, nl H+ la, m) = a n8n.m+l , (a, nl H_la, m) = bn8n,m - l ,

where an and bn are some constants. It follows that a

= (a, nl W21a, n) = (a, nl H+H_ la, n) = (a, nl H+ la, n - 1) (a,n - 11 H_la, n) = anbllo

If we assume that the representation is unitary, then all Wj'S will be hermitian, (H+)t = H_, so an = b~ and a = lanl2 2: O.

876

27. LIE GROUPS AND LIE ALGEBRAS

If ex = 0, then an = 0 and bn = 0 for all n, Consequently, H+ = 0 = H_, i.e., there are no raising or lowering operators. It follows that there are only two spin states, corresponding to the maximum and the minimum eigenvalues of Ho.

helicity of massless particles

A natural axis for the projection of spin is the direction of motion of the particle. Then the projection of spin is called helicity. We summarize our discussion in the following theorem. 27.5.12. Theorem. In the absence ofany interactions, a massless relativistic particle is specified by its spin and its helicity. The former taking on integer or halfodd-integer values s, the latter having values +s and -s. Theorems 27.5 .11 and 27.5.12 are beautiful examples ofthe fruitfulness of the interplay between mathematics and physics. Physics has provided mathematics with a group, the Poincare group, and mathematics, through its theory of group representation, has provided physics with the deep result that all particles must have a spin that takes on a specific value, and none other; that massive particles are allowed to have 2s + I different values for the projection of their spin; and that massless particles are allowed to have only two values for their spin projection. Such far-reaching resnlts that are both universal and specific makes physics unique among all other sciences. It also provides impetus for the development of mathematics as the only dialect through which nature seems to communicate to us her deepest secrets. Ifex > 0, then the resnlting representations will have continuous spin variables. Such representations do not correspond to particles found in nature; therefore, we shall not pursue them any further.

27.6 Problems 27.1. Show that the set G =

GL(n,~)

x

~n

equipped with the "product"

(A, u)(B, v) sa (AB, Av + u) affine group

forms a group. This is called the affine group. 27.2. Show that m : U x U --> ~ defined in Example 27.1.5 is a local Lie group. 27.3. Find the multiplication law for the groups in (b) and (c) of Example 27.1.8. 27.4. Show that the one-dimensional projective group ofExample 27.1.8 satisfies all the group properties. In particnlar, find the identity and the inverse ofan element in the group. 27.5. Let G be a Lie group. Let S be a subgroup of G that is also a submanifold of G. Show that S is a Lie group. 27.6. Show that thc differential map of v : GL(V) --> IHI, defined by lfr(A) = AAt, where IHI is the set of hermitian operators on V, is surjective. Derive Equation (27.11).

27.6 PROBLEMS 27.7. Verify that 19 sa Rg'

0

877

L g is an isomorphism.

27.8. Prove Proposition 27.1.21. 27.9. Start with Equation (27.19) and use the fact that second derivative is independent of the order of differentiation to obtain

Now use the chain rule aUiKjaa" = (aUiKjaXj}(aXj jaa,,) and Equation (27.19) to get

or (27.76) where

Substituting Equation (27.76) in Equation (27.21) leads to (27.22). Now differentiate both sides of Equation (27.76) withrespectto ap to get ac~u -a-UiK =0. up

With the assumption that the UiK are linearly independent, conclude that the structure "constants" areindeedconstants. 27.10. Find the invariant Haar measure of the general linear group in two dimen-

sions. 27.11. Show that the invariant Haar measure for a compact group satisfies dp g = dprl. Hint: Define a measure vby dVg == dprl andshowthatvisleft-invariant. Now use the uniqueness of the left-invariant Haar measure for compact groups. 27.12. Show that O(p, n - p} is a group. Use this and the fact that '7- 1 = '7 to show that A'7A' = '7. 27.13. Show that the orthogonal group O(p, n - p} has dimension n(n - l}j2. Hint: Look at its algebra o(p, n - pl.

878

27. LIE GROUPS AND LIE ALGEBRAS

27.14. Let x = (Xj, X2, x3, xo) be a timelike or (null) 4-vector withxo > O. LetA be a proper orthocbronous transformation. Show that x' = Axis also timelike (null). Hint: Consider the zeroth component of x' as an inner product of (Xj, X2, X3, xo) and another vector and use Schwarz inequality. 27.15. Starting with the definition of each matrix, derive Equation (27.40). 27.16. Let OJ and 02 be derivations of a Lie algebra u, Show that Oj02 es OJ 002 is not a derivation, but [OJ, 02] is. 27.17. Let u be a Lie algebra. Verify that adx is a derivatiou of Ufor any X E U, and that adrx,Yl = [adx, adv] 27.18. Show that 1fr : U -+ aD. given by 1fr(X) = adx is (a) a homomorphism, (b) ker 1fr is the center of n, and (c) aD. is an ideal of :De27.19. Suppose that a Lie algebra Ucan be decomposed into a direct sum of Lie subalgebras. Show that each subalgebra is necessarily an ideal of e. 27.20. Show that if 1fr is an automorphism of n, then ad,,(x) =

1fr 0

adx

0

1fr- j

VXEU.

Hint: Apply both sides to an arbitrary element of e.

27.21. Show that for any Lie algebra,

is completely antisymmetric in all its indices.

27.22. Show that the Killing form of Uis invariant under all automorphisms of n, 27.23. Show that the translation generators P j ofthe poincare algebra p(p, n - p) form a commutative ideal. 27.24. Find the Carlan metrics for 0(3, 1) and p(3, 1), and show directly that the first is semisimple but the second is not. 27.25. Show that the operation on a compact group defined by (ulv)

==

fa

{TguITgv)dJLg

is aninner product. 27.26. Show that the Weyl operator Ku is hermitian. 27.27. Derive Equations (27.53) and (27.54). Hint: Follow the finite-group analogy.

27.6 PROBLEMS

879

27.28. Suppose that a Lie group G acts on a Euclidean space JRn as well as on the transform as the ithrow of space of (square-integrable) functions ,c(JRn). Let

¢!.)

the oth irreducible representation. Verify that the relation na

Tg¢i(·)(X) = LTj~·)(g)¢j·)(x. g-I) j~1

defines a representation of G.

27.29. Show that G L(V) is not a compact group. Hint: Find a continuous function G L (V) --> C whose image is not compact. 27.30. Suppose that T : G --> G L(V) is a representation, and let v®r ==V<8l ... <8lV r times

be the r-fold tensor product ofV. Show that T®r : G --> GL(v®r), given by T~r (VI, ... , v r ) = Tg(VI) <8l ... <8l Tg (v r),

is also a representation.

27.31. Suppose that in Example 27.4.2, we set kl = 2 for our treatment of n = 2, r = 3. Show that Yz(ek, <8l ek, <8l ek,) does not produce any new vector beyond what we obtained for kl = 1. 27.32. Show that gij g" Ciks is anti symmetric in j and r, 27.33. Operate L+ on 100) = L:~~-l C(ll; Dim, -m; 0) 11m; I, -m) and use L+ 100) = 0 to find a recursive relation among C(ll; Dim, -m; 0). Use normalization and the convention that C(ll; Dim, -m; 0) > 0 to show that C(ll; Dim, -m; 0) = (-li-m / '/21 + I (see Section 12.3). 27.34. Show that the generators of so(3, I), M == (MI, Mz, M3) == (MZ3, M31, MIZ), N == (NI,

Ns, N3)

ea (MOl, Moz, MQ3),

satisfy the commutation relations [Mi, Mjl

= -EijkMk,

[Ni, Njl

= EijkMko

[Mi, Njl

= -EijkNk,

and that M Z - N Z and M . N commute with all the M's and the N's.

27.35. Let the double-indexed "metric" of the Poincare algebra be defined as rs

mn

gij,kl = Cij,mnCkl,rs

+ Cij,mGkl,r' r m

where the structure constants are given in Equation (27.69). Show that gij,kl = 2(n - I)(ryjkryi/ -

nonn).

880

27. LIE GROUPS AND LIE ALGEBRAS

27.36. Show that [M2, Mijj = 0, and [M 2 , Pk] = 4MkjPj

+ 2(n -

I)Pk·

27.37. Show that the vector operator Ci

== MijPJ. = nkiJMijPk

satisfies the following commutation relations:

Show also that [C2 , Mjk] piCi = -(n _I)P2 ,

= 0, CiPi = 0, and [C2, Pil = {2Ci + (n -I)P;Jp2 .

27.38. Derive Equation (27.72) and show that Pj's.

wi , ...i"-3

commutes with all the

27.39. Let ex sa (Xl, ... , x n) be any unit vector in JRn. (a) Show that a matrix is '1/-orthogoual, i.e., it satisfies Equation (27.33), if and only if its colunms are '1/-orthogoual. (b) Show that there exists an A E O(p, n - p) such that ex = Ael where e\ = (1,0, ... ,0). Hint: Find the first colunm of A and use (a). (c) Couclude that 0 (p, n - p) is transitive in its action on the collection of all vectors of the same length. 27.40. Verify directly that when the 4-momentum has only a time component, the Casimir operator W2 = W . W reduces essentially to the total angular momentum operator. 27.41. Verify that for the case of a massless particle, when Po = (p, 0, 0, p), Wo = WI = M23,

W2 = 2p(M13 - M03),

W3 = 2p(M02 - M12),

and that W2 = W· W annihilates l1fJpo ) '

Additional Reading 1. Barut, A. and Raczka, R Theory ofGroup Representations andApplications, World Scientific, 1986. A comprehensive introduction to Lie groups and Lie algebras using the modem language of manifolds. lutended for physicists and mathematicians alike, the first chapter is a long introduction to Lie algebras. 2. Chevalley, C. Theory ofLie Groups, Princeton University Press, 1946. Still a relevant classic for Lie groups and manifold theory written by one of the major contributors to the subject.

27.6 PROBLEMS

881

3. Gilmore, R. Lie Groups,Lie Algebras, and Some ofTheir Applications, Wiley, 1974. Although the concept of manifold is introduced, no heavy use of it is made in the treatment ofLie groups and Lie algebras, and the "parametric" method of Sophus Lie is employed throughout the book. Nevertheless, the book does a good job of classifying the Lie algebras for outsiders. 4. Hamermesh, M. Group Theory and Its Application to Physical Problems, Dover, 1989. Does not use the modem language of manifolds, but does a good job of introducing Lie groups and Lie algebras via "parameters."

28

_

Differential Geometry

The elegance of the geometrical expression of physical ideas has attracted mnch attention ever since Einstein proposed his geometrical theory of gravity in 1916. Such an interpretation was, however, confined to the general theory of relativity until the 1970s when the language of geometry was found to be most suitable, not only for gravity, but also for the other three fundamental forces of nature. Geometry, in the form of gauge field theories of electroweak and strong interactions, has been successful not only in creating a model-the so-called standard model-that explains all experimental results to remarkable accuracy, but also in providing a common language for describing all fundamental forces of nature, and with that a hope for unifying these forces into a single all-embracing force. This hope is encouraged by the successful unification of electromagnetism with the weak nuclear force through the medimn of geometry and gauge field theory. The word "geometry" is normally used in the mathematics literature for a manifold on which a "machine" is defined with the property that it gives anmnber when two vectors are fed into it. Symplectic geometry's machine was a nondegenerate 2-form. Riemannian (or pseudo-Riemannian) geometry has a symmetric bilinear form (metric, inner product). Both of these geometries are important: Symplectic geometry is the natural setring for Hamiltonian dynamics, and Riemannian geometry is the basis of the general theory of relativity. In this chapter, we shall study the latter. 1 However, before introducing the metric, let us investigate some related structures that are independent of a metric.

IThischapter reallybelongsto theprevious part of the book; however, becauseof ouruse of certain Lie-group theoretic ideas in Sections28.4 and28.6, we haveincluded it here.

28.1 VECTOR FIELDS AND CURVATURE

883

28.1 Vector Fields and Curvature A manifold is, in general, not flat (flatness will be defined later). One way to "feel" the curvatnre of a space intrinsically is to translate a vector parallel tn itself along different paths and compare the final vectors. In a flat space the two vectors at the end will be the same, but not in a general space. An illnstration is provided by the surface of a sphere. Assume that we have a vector perpendicular to the equator. To exaggerate the effect of curvatnre, we move the vector parallel to itself on the equator a qnarter of the way around the sphere, then all the way to the north pole. Alternatively, we start with the vector again perpendicular to the equatnr, but this time we move it parallel to itself directly to the north pole. Clearly, the two fiual vectors will uot be the same; in fact, they will be perpeudicular to oue another. The above intnitive discussion should help to make it clear that to find the curvatnre of space, we look at how vectors change. In analogy with the exterior derivative and forms, we want to introduce a derivative that operates on vectors. In fact, since we already have the exterior derivative available, let us see if we can extend it to vectors. Consider an arbitrary I-form w and an arbitrary vector field v. The pairing (w, v) is a real-valued function f ou which we know how the exterior derivative acts. A uatnral extension of d (which we denote by the same symbol) is given as follows: df = d (w, v) sa (dw, v)' + (w, dv)' ,

vector-valued Horms

(28.1)

where we have used a prime to designate the new pairing. In general, this new pairing may be different from the ordinary pairing, because for the participants in the latter, no exterior differentiation is defined. As we shall see below, we indeed have to change the old pairing slightly for the new pairing to make sense. The LHS of Equation (28.1) is a I-form. The first term on the RHS is a 2-form coutracted with a vector, i.e., a l-form, For the secoud term to be a l-form, dv mnst be a tensor that contracts with a l-form to give a l-form, We say that dv is a vector-valued I -form, Let us take a basis lei) and its dual {€j) and express Equation (28.1) in terms of components in thosebases.Then, on the one hand, with w = Wi ei and v = v j eiwe have (w, v) = WiVi and df = d (w, v) = (dWi)Vi

+ Widvi.

(28.2)

On the other hand, . df = (dWi A .

. €'

.., .. . , + Wid€', vJej) + (Wi€', dvJej + vJdej) . , .

.

,

..

I

..

=v J (dWi A€',ej) +VJWi (d€',ej) +Wi (€',dvJej) +WiVJ (€',dej) = v j (dWi A €i, ej)' .

,

+ Widvj (E', ej) +VjWi {(d€i, ej)' + (€i, dej}'j, '--:-' ISj

(28.3)

884

28. DIFFERENTIAL GEOMETRY

where we have assumed that the primed pairing is equivalent to the old pairing when no derivatives are inside. If Equations (28.2) and (28.3) are to hold for arbitrary w and v, we must bave? .

I

0= ide", ej)

..

.

+ (E', dej)

.,

,

,

(dWi)V' = vl (dWi A E', ej) .

(28.4)

The first relation is simply the fact that the exterior derivative of 8~ = (Ei , e j) is zero. The second relation defines the order of contraction of vectors and higherorder forms in the new pairing. Since dWj is a I-form, we can write it as dWj = Uikf.

k,

so that

do» /\



i

k i = aik€. /\ e ,

where

aik = -aki·

Then, the second equation of (28.4) gives

This equation demands that the new pairing of a vector with a wedge product of l-forms be defined as _ (€ k Aei ,ej )' =

. (k e Aei)

-lej

=

, k) , -21 ('Okj€i -Oi}e

where iej is the interior product of Definition 26.5.8. We summarize the properties of the new pairing in the following equation:

. (dE',ej)

,

.

+ (E',dej)

,

=0,

k i )' =-lej€Ae . (k i) =-zUkje-Uij€. 1 (' i , k) (eAe,ej

(28.5)

In differential structures everything takes place locally, and translations and movements are all infinitesimal. Let us look at the exterior derivative from this point of view. For a real-valuedfimction on M, we have

af d 2 f =d ( -.dx ax'

2 i) = I: (-.-.-, a f ..

l<'

axlax'

2

f.) dx i Adx. i -a. ax'ax l

Thus, d 2 f = 0 means that the mixed partial derivatives are independent of the order of differentiation-a familiar result. Geometrically, this means that for small displacements, dx i and dx j , the value of a function is the same if one moves in two (perpendicular) directions, once in a given order and then in reverse order. This is 2Theseconsistency relations applyto the first application of exterior derivative to thepairing. Higher-order applications, or, equivalently, application of d toa pairing involving wedgeproducts of forms, mayrequire newconsistency relations. Fortunately, for ourpurposes, the first-order consistencyrelations will be sufficient.

28.1 VECTOR FIELDS AND CURVATURE

885

true even if the space is cnrved. For flat spaces, we know that the same conclusion holds for displacement (parallel to themselves) of vectors. When we interpret d 2v as the change in a vector as it is displaced in two different directions, then the example of the sphere above suggests that d 2v must be related to curvature. Let us find this relation. Starting with a basis [e, } and an arbitrary vector v = vi er, operate on it with d twice, keeping in mind that its action on functions and differential forms is exactly the same as the exterior derivative defined before: dv = dviei + videi and

d 2v = d 2vi ei + (_I)ldv i A de, '-.-'

=0

":0

+ dv i A de,.+vid 2ei =

vid2ei'

(28.6)

This equation has a remarkable property: It leaves the components of v undifferentiated! In other words, regardless of how any given two vectors v and w vary away from the point P of the manifold, d 2 v = d 2w as long as the two vectors are equal at P. More importantly, if we take the linearity of d (and therefore, d 2 ) into account, we obtain (28.7)

It appears that d 2 v depends not on external objects (vectors), but on the intrinsic property of the manifold, i.e., how it "curves" away from P. To find this curvature, expand the vector-valued I-form de, as (28.8)

extracting the curvature 2-form from second exterior derivative of basis veetors

As we shall see shortly, one has to be cautious to know which index is raised or lowered. In the formulas above, this caution has been observed by leaving blank the original position of the raised (or lowered) index. Differentiating Equation (28.8) once more, we obtain the vector-valued 2-form

d2ei = de] A Wii + ei 181 dWii = (ek 181 wki) AWii + ei 181 dWii _",,(k ""di_""(di+i k) =ek'<::JJ Wj!\W i) +ej'Ol wi-ej'O' Wi WkAWi· i

(28.9) The expression in parentheses is a 2-form, called the curvature two-form: or

(28.10)

With this notation, Equation (28.6) becomes .o, Oi i "" e v, i -o, Oii=ej'Ol d 2v=v id2 ei=vej'Ol iV =e'O'

where the last expression is simply an abbreviation.

(28.11)

886

28. DIFFERENTIAL GEOMETRY

So far, we have been dealing with the exterior derivatives of the basis vectors {e.}, What about the exterior derivatives of the dual basis vectors {Ei}? They are closely related to {de;}, as the following argument shows. The first relation of (28.5), as well as Equation (28.8) and the fact that Ei pairs up with ek yield . , . , k . I k . j (ei, de') = - (dei, El) = - (ek I8i W i' El) = -w i (eko El) = -w i

(28.12) On the other hand, since dEi is a 2-form, it can be written as

where

i _ Y km -

-

yi

mk

and

where we used the second equation in (28.5). Comparing (28.13) with Equation (28.12) yields a{ = -Wi i . We therefore have (28.14) Equations (28.8) and (28.14) show that the {wij} give all the information about how the bases {ei} and {Ei} change with infinitesimal movement away from a point

P. 28.1.1. Box. If we can find the {wij}, we will know the (local) geometry of the manifold.

connection coefficients

From the definition of r ii k in Equation (28.8), we have wii = riikEk. The functions riik are called the connection coefficients. Because of (28.8), these coefficients are antisymmetric in their first two indices. On the other hand, (28.14) gives

(28.15)

If coordinate frames are used for basis vectors, so that

Ek

= dx"; the LHS of (28.15) will be zero and the coefficients on the RHS of Equation (28.15) must vanish, i.e., the connection coefficients are synunetric in their last two indices:

in coordinateframes.

I

(28.16)

28.2 RIEMANNIAN MANIFOLDS

887

Equation (28.15) shows further that by calculating de k we merely determine the anti symmetric combination rijrn - rimj. However, the antisymmetry orrijm in its first two indices can be used to determine it completely. Let Cijm ea rijm - rim} be the coefficients that can be read off from (28.15). Then one can show that

(28.17) Once the P's are determined by this relation, they can be used to write W}i'S as a linear combination of the dual basis vectors.

28.2 Riemannian Manifolds As mentioned before, manifolds that possess a metric are important. The general theory of relativity, for example, is entirely based on the existence of a metric. In fact, it is the job of that theory to determine the metric of 4-dimensional space-time from a knowledge of the distribution of matter. Riemannian and pseudo-Riemannian manifolds defined

orthonormal frames

28.2.1. Definition. A Riemannian manifold is a differentioble manifold M with a symmetric tensor field 9 E T~(M), called the metric, such that at each point P E M, gjp is apositive definite inner product. A manifold with an indefinite inner product at each point is called a pseudo-Riemannian manifold. With 9 defined on M, we can obtain orthonormal vectors at each point of M. That is, we can constroct orthonormal frames {ei} such that

at each point P EM.

28.2.1 Curvature via Connection We can find a relation between the metric tensor and Wi} by taking the exterior derivative of both sides of gil = ei . e} and using Equation (28.8): dgi}

= (dei) . e} + e, . (de}) = (ek ® wk;J. e) + ei . (ek ® w k) sa w\ ® (ek' e})

structure equations, integrability condltlon, and curvature matrix

+ J} ® (e, . ek) = wkigk} + Wk}gik '" W}i + wij.

In patticniar, if we work in an orthonormal basis, then gil = ±8ij, and the LHS will be zero' In such a case, we obtain the following antisymmetry condition for the I-forms Wi}: Wij+Wji

=0.

30rthonormality is not really required. All that is necessary is for gjj to be constant.

(28.18)

888

28. DIFFERENTIAL GEDMETRY

We now develop an algoritlnn to determine the local curvature of the manifold. Choose an orthonormal basis and its dual and introduce the matrices

e = (el ... em),

e=C),

0=

0

W12

W13

Wl m

-W12 -w13

0

W23

W2m

-W23

0

W3m

-Wl m

-W2m

-W3m

0

whose elements are one-forms, or vectors. Write Equations (28.8), (28.18), (28.14), and (28.9) as

de= eGO, e=ee,

d

2

0+0' =0,

where

de e

= -GOAe;

=dO + OA (GO).

(28.19)

The matrix G is the matrix of the metric with components gij = ±8ij. It is introduced to raise the indices when the equations are wrilten in component form," The first two equations in (28.19) are called the structure equations, the third is called the integrability condition, and e is called the curvature matrix.

No great mindof thepasthas exerted a deeperinfluence on the mathematics of the twentieth century than Georg Friedrich

Bernhard Riemann (1826-1866), the son of a poor country minister in northern Germany. He studied the works of Euler and Legendre while he was still in secondary school, andit is saidthat he mastered Legendre's treatise on thetheory of numbersin less than a week. Buthe was shy andmodest, withlittle awareness of his ownextraordinary abilities, so attheage of 19 he wentto theUniversity of G5ttingen withtheaimof pleasing his father by studying theology andbecominga minister himself. Fortunately, this worthy purpose soon stuckin his throat, andwithhis father's willing permissionhe switchedto mathematics. The presence of the legendary Gauss automatically made Gdttingen the center of the mathematical world. ButGauss wasremote andunapproachable-particularly to beginning students-and after only a yearRiemann left thisunsatisfying environment andwentto the University of Berlin. There he attracted the friendly interest of Dirichlet and Jacobi, and learned a great deal from both men. Tho years laterhe returned to Gottingen, where he obtained his doctor's degree in 1851. During the next 8 years, despitedebilitating poverty, he created his greatest works. In 1854 he was appointed Privatdozent (unpaid lecturer), whichatthattimewas thenecessary first stepon the academic ladder. Gauss died in 1855, and Dirichlet was called to Gottingen as his successor. Dirichlet helpedRiemann in every

wayhe could,firstwitha smallsalary(aboutone-tenth of that paid to a full professor) and thenwitha promotion to anassistant professorship. In 1859 he also died,andRiemann was "Stnctty speaking, we shoulduse G- 1 instead of G.But since gjj

= ±8ij, thetwo areidentical.

28.2 RIEMANNIAN MANIFOLDS

889

appointed as a full professor to replace him. Riemann's yearsof poverty were over, buthis health was broken. At the age of 39 he died of tuberculosis in Italy, on thelast of several trips he undertook in order to escape the cold, wet climate of northern Germany. Riemann had a short life and poblishedcomparatively little, but his works permanently altered the courseof mathematics in analysis,geometry, andnumber theory. It is said that the three greatest mathematicians of modem times are Euler, Gauss, and Riemann. It is a curiosity of nature thatthese three names are among the most frequently mentioned namesin thephysicsliterature as well. Asidefromtheindirect use of his namein theapplication of complexanalysis inphysics,Riemannian geometryhasbecomethemost essential building block of all theories of fundamental interactions, starting with gravity, which Einstein formulated in this language in 1916. As part of the requirement to become a Privatdozent, Riemann had to write a probationary essay and to present a trial lecture to the faculty. It was the custom for the candidate to offerthree titles, andthe headof his department usually accepted the first. However, Riemann rashly listed as his third topic the foundations of geometry. Gauss, who had been turning this subject over in his mind for 60 years, was naturally curious to see how thisparticular candidate's "gloriously fertile originality" wouldcope withsucha challenge, andto Riemann's dismayhe designated this as thesubject of thelecture. Riemann quickly torehimself awayfrom his otherinterests at the time-"my investigations of the connection betweenelectricity, magnetism, light,and gravitation"-andwrote his lecture in thenexttwo months. Theresultwas one of thegreat classical masterpieces of mathematics, andprobably the most important scientific lecture evergiven. It is recorded thateven Gauss was surprised andenthusiastic.

Bianchi identity

We can derive further integrability conditions. For instance, applying d to the third equation of (28.19) gives

0= d 2e

= -d(GO) /\ e + GO /\ de = -GdO /\ e -

(GO) /\ (GO) /\ e.

Multiplying both sides on the left by G = G- 1, we obtain

0= -[dO + 0 /\ (GO)] /\ e = -El /\ e.

(28.20)

Similarly, the reader may show that dEl = 0/\ (GEl) - El /\ (GO).

Riemann curvature tensor

(28.21)

This is called the Bianchi identity. The i jth element of the matrix of El can be written as (Jij =

!Rijkl,} /\ el ,

(28.22)

which defines the components Riju of the Riemann curvature tensor. The antisyrnrnetry of the matrix El (showing this is left as a problem for the reader) and Equation (28.22) give Rijkl

+ R jikl

= 0,

Rijkl

+ Rijlk = O.

(28.23)

890

28. DIFFERENTIAL GEOMETRY

Square brackets mean antisymmetrization.

Similarly, the relation El Ae = 0 ofEqnation (28.20) can be shown to be eqnivalent to Ruu

+ Riklj + Riljk =

0

Ri[jkl] = 0,

and

(28.24)

where 28.2.2. Box. The enclosure of indices in square brackets means complete antisymmetrization of those indices.

When coordinate bases are used, Wij is no longer antisymmetric, but we still have dgij = Wij

+ Wji

es rijkdxk

+ F jikdxk.

Since dgij = (agij/axk)dx k, we get agij glj,k '" ax k = rijk

Christoffel symbol

+ F jik-

Using Eqnations (28.16) and (28.25), we can readily show that 1

rijk = 'J.(gij,k + gik,j - gkj,i) '"

connection between infinitesimal displacement, arc length, andmetric tensor

(28.25)

I (agij

2:

axk

+

agik agkj) ax j - axi .

(28.26)

This is the Christoffel symbol nsed in classical tensor analysis. Now we consider the connection between an infinitesimal displacement and a metric. Let P be a point of M. Let y be a curve through P such that y (c) = P. For an infinitesimal number Su, let pI = y(c + 8u) be a point on y close to P. Since the xi are well-behaved functions, xi (PI) _xi (P) are infinitesimal real numbers. Let;i = xi (PI) _xi (P), and construct the vector v = ;i ai, where {ai} consists of tangent vectors at P. We call v the infinitesimal displacement at P. The length of this vector, g(v, v), is shown to be gij;i;j. This is called the arc length from P to r, and is naturally written as ds 2 = gij;i; j . It is customary to write dx i (not a I-foun!) in place of

;i: (28.27)

where the dx' are infinitesimal real numbers.

Elwin Bruno Christoffel (1829--1900) came from a famity in the cloth trade. He attended an elementary school in Montjoie (whichwasrenamedMonschau in 1918)but thenspenta number of years being tutored at home in languages, mathematics, and classics. He attended secondary schools from 1844 unti11849. At first he studied at the Jestdt gymnasium in

28.2 RIEMANNIAN MANIFOLDS

891

Cologne hutmovedtothe Friedrich-Wilhelms Gymnasiumin the sametownfor at leastthe three final years of his school education. He was awarded the final school certificate with a distinction in 1849. The next year he went to the University of Berlin and studied under a number of distinguished mathematicians, including Dirichlet. After one year of military service in the Guards Artillery Brigade, he returned to Berlin to study for his doctorate, which was awarded in 1856 with a dissertation on the motion of electricity in homogeneous bodies. His examiners included mathematicians and physicists, Kummer being one of the mathematics examiners.

At this point Christoffel spent three years outside the academic world. He returned to Montjoie, where his mother was in poor health, but read widely from the works of Dirichlet. Riemann, and Cauchy. It has been suggested that this period of academic isolation had a major effect on his personality and on his independent approach towards mathematics. It was during this time that he published his first two papers on numerical integration, in 1858, in which he generalized Gauss's method of quadrature and expressed the polynomials that are involved as a determinant, This is now called Christoffel's theorem. In 1859 Christoffel took the qualifying examinatiouto become a university teacher and was appointed a lecturer at the University of Berlin. Four years later, he was appointed to a chair at the Polytechnicum in Zurich, filling the post left vacant when Dedekind went to Brunswick. Christoffel was to have a huge infiuence on mathematics at the Polytechnicum, setting up an institute for mathematics and the natural sciences there. In 1868 Christoffel was offered the chair of mathematics at the Gewerbsakademie in Berlin, which is now the University of Technology of Berlin. However, after three years at the Gewerbsakademie in Berlin, Christoffel moved to the University of Strasbourg as the chair of mathematics, a post he held until he was forced to retire due to ill health in 1892. Some of Christoffel's early work was on conformal mappings of a simply connected region bounded by polygons onto a circle. He also wrote important papers that contributed to the development of the tensor calculus of Gregorio Ricci-Curbastro and Tullio Levi-Clvlta. The Christoffel symbols that he introduced are fundamental in the study of tensor analysis. The Christoffel reduction theorem, so named by Klein,solves the local equivalence problem for two quadratic differential forms. The procedure Christoffel employed in his solution of the equivalence problem is what Ricci later called covariant differentiation; Christoffel also used the latter concept to define the basic Riemann-Christoffel curvature tensor. His approach allowed Ricci and Levi-Civita to develop a coordinate-free differential calculus which Einstein, with the help of Grossmann, turned into the tensor analysis, the mathematical foundation of general relativity.

In applications, it is common to start with the metric tensor 9 given in terms of coordinate differential forms:

(28.28) Then the orthonormal bases {eil and {ei } are constructed in terms of {ail and {dx i }, respectively, and are utilized as illustrated in the following examples. The

892

28. DIFFERENTIAL GEDMETRY

equivalence of the arc leugth [Equation (28.27)] and the metric [Equation (28.28)] is the reason why it is the arc length that is given in most practical problems. Once the arc length is known, the metric gij can be read off, and all the relevant geometric quantities can be calculated from it.

28.2.3. Example. Let us look at a few examples of arc lengths and the corresponding metrics. (a) For ds 2 = dx 2 + di + dz 2, 9 is the Euclideanmetric of]R3, with gij = 8ij' (b) For ds 2 = -dx 2 - dy2 - dz 2 + dt 2, 9 is the Minkowski (or Lorentz)metric of ]R4,

Friedmann metric

with gij = 'fJij, where 7Jxx = 'fJyy = '11zz = -YIn = -1 and TJij = 0 for i =1= j. (c) For ds 2 = dr 2 + r 2(d0 2 + sin20dq>2), the metric is the Euclidean metric given in spherical coordinates in IR3 with grr = I, gee = r 2 , g(/l({J = r 2sin2 e. and all other components zero. (d) For ds 2 = a 2d02 + a 2 sin2 Odq>2, the metric is that of a two-dimensional spberical surface, with see = a 2 , gtptp = a 2 sin2 e, andall other components zero. (e) For

Schwarzschild metric

the metric is theFrledmann metric used in cosmology. Here gtt = 1, sx« = -[a(t)]2, g99 = -[a(t)]2 sin2 x. g~~ = _[a(t)]2 sin2 X sin2 0, and all other components are zero. (f) For

the metric is the Sehwarzsehild metric with Btt = 1- 2M/r, grr = -(I - 2M/r)-t, _r 2 , g(fl'P = _r2 8in2 9, andall other components zero. For eachof thearc lengths above, we have an orthonormal basisof one-forms: (a) 9 =.1 ®.I +.2 ®.2 +.3 ®.3 with. 1 = dx,.2 = dy,.3 = dz: (b) 9 = _.1 ®.I _ .2 ®.2 _ .3 ®.3 +.0 ®.O with. 1 = dx, e2 = dy, .3 = dz, .0 = dt; (c) 9 = t!' ®.' +.9 ®.9 +.~ ®.~ with e" = dr,.9 = rdO,.~ = rsinOdq>; (d) 9 = .9 ®.9 +." ®.~ with.e = adO,.~ = a sinOdq>; (e) 9 = .' ®.' _.x ®.x _.9 ®.9 _.~ ®.~ with.' = dt,.X = a(t)dx,.9 = a(t) sinXdO,.~ = a(t) sin XsinOdq>; (f) 9 = .' ® .' - .r ® .r _ .e ® .9 _ .~ ®.~ with.' = (I - 2M/r)I/2dt, .' = (1- 2M/r)-1/2dr,.9 = rdO,.~ = rsinOdq>. III gfJf) =

28.2.4. Example. In thisexample, we examine thecurvilinear coordinates usedin vector analysis. Recall that in terms of these coordinates the displacement is given by ds 2 = hr(dql)2 + h~(dq2)2 + h~(dq3)2. Therefore,the orthonormalone-forms are. 1 = hI dqj, .2 = h2dq2, .3 = h3dq3. We also note (see Problem26.24) that

(28.29)

28.2 RIEMANNIAN MANIFOLDS

893

Weuse this equationto find the Laplacianin terms of Ql, Q2. and q3:

Differentiating once more, we get

(28.30)

Since {e1, e2, e 3} are orthonormal one-forms (as are {dx, dy, dz}), the volume elements e 1 A e 2 A e 3 and dx /\ dy /\ dz are equal. Thus, we substitute the latter for the former in (28.30), compare with (28.29), and conclude that

"12

r-- _1_ [...!.. (h2h3!L) +...!.. (hlh3!L) +...!.. (hlh2 at)] hlh2h3 aql hi aql aq2 h2 aq2 aq3 h3 aq3 '

which is the result obtained in curvilinear vector analysis. 28.2.5. Example. Let M = R 2, and suppose that the arc length is given by ds 2 = (dx 2 + dy2)j y 2 We can write the metric as 9 =.1 18>.1 + .218>.2 if we define I €

dx =Y

and

2 dy e=y

The dual vectors {e1, e2} are orthonormal one-forms and G = 1, so we need not worry about raising and lowering indices. Inspection of the definition of (3.1 and e l, along with the fact that dx(a x ) = dy(a y) = I and dx(a y) = dy(ax ) = 0, immediately gives ej = yax and ej = yay. To find the curvature tensor, we take the exterior derivative of the ei 's:

d. 1

= d (I) ydx = -

I

y2dY rvdx

=. I A. 2 ,

(28.31)

894

28. DIFFERENTIAL GEOMETRY

From these equations, the antisymmetry of the w's, and Equation (28.14), we can read off = W22 = Oandw12 = -W21 = _e 1. Thus, thematrixOis

Wij. They are Wl1

0- (

W 1Z) =

0

o

-W12

(0

1

_e )

o '

e1

whichgives

dO=

C~1

-n -n

and

OAO=

(~1

A

(~

=0.

Therefore, the curvature matrix is 0

e=dO=(e 1 Ae Z This shows thatthe only nonzeroindependent componentof theRiemanncurvature tensor isRl2lZ=-1. l1li

28.2.6. Example. Fora spherical surface of radius a, theelement of length is

ds Z = aZdO Z + a Z sin Z OdrpZ. The orthonormal forms are e(} = ade anderp = a sin Bdep, and we have derp

G = 1,

1 cotBe() a

= a cos 9dfJ A dcp = -

A

e'P.

Thematrix n cannowbe read off:5 0= (

cot~e~ - c~o e } a

A straightforward exterior differentiation yields

0 dO=

1 Aerp ( --e(} aZ

Similarly, 0 A 0 = O. Therefore, the curvature matrix is

e = zn = ~ ( 0 a2

_,/J 1\ erp

e 8

A e~) O

Theonly independent component of theRiemann curvature tensor is R(Jrp(Jtp = 1/ a2 , which is constant, as expected for a spherical surface. II

28.2 RIEMANNIAN MANIFOLOS

What is a flat manifold?

895

It is clear that when the gij in the expression for the line element are all constants for all points in the manifold, then e i will be proportional to dx' and de i = 0, for all i. This immediately tells us that n = 0, arid therefore e = 0; that is, the manifold has no curvature. We call such a manifold flat. Thus, for ds 2 = dx 2 + d y 2 + dz 2 , the space is flat. However, arc lengths of a flat space come in various guises with nontrivial coefficients. Does the curvature matrix recognize the flat arc length, or is it possible to fool it into believing that it is privileged with a curvature when in reality the curvatore is still zero? The following example shows that the curvatore matrix can detect flatoess no matter how disguised the line element is!

e

28.2.7. Example. In spherical coordinates, the tine element (arc length) of the lIat Enclidean space R' is ds 2 = dr 2 + r 2de2 + r 2 8in 2 Bd
expression above: er = dr,

= rde.

eO

e({J

= r sin Bdcp,

G = 1.

Taking the exterior derivatives of these one-forms, we obtain

de" = d 2r = 0, 1\ dO +r ___ d 20 = €r 1\

(€9) = ~€r 1\ €9, r r

d€9

= d(rdO) = dr

d€~

= d(rsinO) I\drp = sinOdr 1\ drp+rcosOdO I\drp

~O

=sin8er

A(~)+rcose(€9)1\( €~O) rsmfJ r rsm

I = _e r Ae({J r

9 + -cotO - e Aef{J. r

We can now use Equation (28.14) to find the matrix of one-forms n. In calculating the elements of fl, we remember that it is a skew-symmetric matrix, so all diagonal elements are zero. We also note that dek = 0 does IWt imply that J'j = O. Keeping these facts in mind, we can easily obtain n (the calculation is left as a problem for the reader):

I 9

--€

r

o cot eerp r

The exterior derivative of this matrix is found to be

.oj

-- E9 Ae T C.otO

[2

_E 9 A

r2

o SNote that

dE(}

= 0 does not imply

W12

= O.

eqJ

'

896

28. DIFFERENTIAL GEOMETRY

whichis precisely (the negative of) theexterior product n A fl, as thereader maywish to verify. Thus, e = dn + nAn = 0, and the spaceis indeedfiat! III In all the foregoing examples, the curvatnre was calculated intrinsically. We neverhad to leave the spaceandgo to a higherdimension to "see" the curvature. For example, in the case of the sphere, the only information we had was the line element in terms of the coordinates on the sphere. We never had to resort to any three-dimensional analysis to discover a globe embedded in the Euclidean IR3. As mentioned earlier, if a spacehasline elements withconstant gij, then theRiemann curvature vanishes trivially. We have also seen examples in which the components of a metric tensor were by no means trivial, but e was smart enough to detect the flatness in disguise. Under what conditions can we choose coordinate systems in terms of which the line elements have gij = ±8ij? To answer this question we need the following lemma (proved in [Flan 89, pp. 135-136]):

28.2.8. Lemma. If 0 is a matrix of i-forms such that dO + 0 /\ (GO) = 0, then there exists an orthogonal matrix A such that dA = AGO. The question raised above is intimately related to the connection between coordinate and orthonormal frames. We have seen the usefulness of both. Coordinate frames, due to the existence of the related coordinate functions, are useful for many analytical calculations, for example in Hamiltonian dynamics. Orthonormal frames are useful because of the simplicity of expressions inherent in all orthonormal vectors. Furthermore, we saw how curvatnre was easily calcnlated once we constructed orthonormal dual frames. Naturally, we would like to have both. Is it possible to construct frames that are both coordinate and orthonormal? The following theorem answers this question:

e

28.2.9. Theorem. Let M be a Riemannian manifold. Then M is flat, i.e., = 0 if and only if there exists a local coordinate system {xi} for which {ail is an orthonormal basis.

Proof The existence of orthonormal coordinate frames implies that {dxi} are orthonormal. Thus, we can use them to find the curvature. But since d(dx i) = 0 for all i, it follows from Equation (28.12) that w i j = OandO = O. So the curvature must vanish. Conversely, suppose that e = O. Then by Lemma 28.2.8, there exists an orthogonal matrix A such that dA = AGO. Now we define the one-form colunm matrix T by T = Ae, where e is the one-form colunm matrix of Equation (28.14) . . Then, using (28.19), we have dT

= d(Ae) = dA A e + Ado = (AGO) /\ 0

-

A(GO /\ 0)

= o.

Thus, dT i = 0 for all i, By Theorem 26.5.14 there must exist zero-forms (functions) xi such that T i = dx', These xi are the coordinates we are after. The basis {ail is obtained using the inverse of A (see the discussion following Proposition 25.1.1). Since A is orthogonal, both {dxi} and {ail are orthononnal bases. D

28.3 COVARIANT DERIVATIVE AND GEODESICS

897

28.3 Covariant Derivative and Geodesics The essence of all geometries are straight lines. The familiar Euclidean geometry is developed entirely based on a number of postulates concerning certain attributes and properties for straight lines. From the physical standpoint, straight lines are those trajectories on which "free" particles-including light-travel. If geometry is the basis of physical theories (general theory of relativity, and, to a lesser degree, electromagnetism and the nuclear interactions), knowledge of straight lines will be crucial.

28.3.1 Covariant Derivative Straight lines are characterized by the "least amount of bending." In flat space, this involves zero bending; but if space is curved, the bending cannot be eliminated. The bending of space is gauged by a test vector as it moves (infinitesimally) along some trajectory. The infinitesimal character of any trajectory is encapsulated in the vector tangent to it at the point of interest. Thus the concept of straight line is tied to the way one vector (the test vector) changes along a second vector (the tangent vector). covariant derivative ofa vector

28.3.1. Definition. Let u, v E X(M) be vector fields on M. The covariant derivative ofv with respect to (or along) u, denoted by Vuv, is defined as Vuv = (dv, u).

(28.32)

So, Vu : X(M) -+ X(M), i.e., Vu maps vector fields to vector fields. Certain properties of the covariant derivative follow from this definition. In most books on differential geometry and relativity, these properties are used to define the covariant derivative. We collect these properties in the following theorem.

Properties of covariant derivative

28.3.2. Theorem. Let u, v, and w be vector fields, and f and h functions on a manifold M. Then the covariant derivative has the following properties: 1. Vuv - Vvu = [u, v]. 2. Vu(fv) = u(f)v + fVuv.

4. Vju+hw(V) = fVuv+ hVwv.

Proof We shall prove the first property, which happens to be the hardest. The rest are easy consequences of the definition and the linearity of pairing. To prove the first property, choose a basis" {e.} and its dual {Ei} and write v = viei' u = uiei, 6Eventually, we shall takethe basis to be a coordinate frame. But for notational convenience, we first work in a general basis.

898

28. DIFFERENTIAL GEOMETRY

and

Then we have

= (dv i, n) ei + vi (dei, u) = [u(vi)]ei + viu j (dei, ej) , (du, v) = (du i, v) ei + u i (de;, v) = [v(ui)]ei + u i v j (dei, ej) ,

(dv, n)

and (dv, u) - (du, v) = [u(v i) - v(ui)]ei

+ viu j

«dei, ej) - (dej, ei»'

(28.33) Let us evaluate the term in parentheses. Using Equation (28.8), we have (dei, ej) - (dej, ei) = (ek ® d'i' ej) - (ek ® Wkj,ei) k = ek (w~, ej ) - ek (w kj , e.) = ek (r~j - r j i).

Thebasischosenabove wasgeneral. However, we arefreetochooseanyconvenient basis to prove a vector or tensor identity. If we choose the basis to be a coordinate frame, then by Equation (28.16), r~j - r kj i = O. Furthermore, [u(v i ) - v(ui)]ei = [u, v]

in a coordinate frame.

Substitutiog these two relations in Equation (28.33), we obtain the first part of the theorem. D

Tullio Levi-Civila (1873-1941), the son of Giacomo LeviCivita, a lawyer who from 1908 was a senator. was an out-

standingstudentat the liceoin Padna.10 1890 he enrolledin the Faculty of Mathematics of the University of Padua. Giuseppe Veronese and Gregorio Ricci-Curbastro were amonghis teachers.

He receivedhis diplomain 1894 and in 1895 became resident professor attheteachers' college annexed tothe Faculty of Science at Pavia. From 1897 to 1918 Levi-Civita taught rational mechanics attheUniversity ofPadua. Hisyears in Padua (where in 1914he married apupil, Libera Trevisani) werescientifically themostfruitful of hiscareer. In 1918he became professor of higher analysis atRome and, in 1920, of rational mechanics. In 1938, struck by the fascist racial laws against Jews,he was forced to give up teaching. The breadth of his scientific interests, his scruples regarding the fulfillment of his academic responsibilities, and his affection for youngpeople made Levi-Civita the leader of a flourishing schoolof mathematicians.

28.3 COVARIANT DERIVATIVE AND GEODESICS

899

Levi-Civita's approximately 200 memoirs in pureand applied mathematics deal with analytical mechanics, celestial mechanics, hydrodynamics, elasticity, electromagnetism, andatomicphysics. His most important contribution to science was rooted in thememoir "Sulletrasformazioni delle equazioni dinamiche" (1896), whichwas characterized by the use of themethods of absolute differential calculusthatRicci hadappliedonly to differential geometry. Inthe"Methodes de calcul differentiel absolus et leurs applications," written withRicci andpublished in 1900 in Mathematische Annalen,there is a complete exposition of the new calculus, whichconsists of a particular algorithm designedto expressgeometric and physical laws in Euclidean andnon-Euclidean spaces, particularly in Riemannian curved spaces. The memoir concerns a very general but laborious type of calculus that made it possibleto dealwithmanydifficult problems. including, according to Einstein, the formulation of the general theory of relativity. Although Levi-Civita had expressed certain reservations concerning relativity in the first years after its formulation (1905), he gradually came to accept the new views. His own original research culminated in 1917 in the introduction of the concept of parallel transport in curved spaces. With this new concept, absolute differential calculus, having absorbed other techniques. became tensor calculus, now the essential instrument of the unitary relativistic theories of gravitation andelectromagnetism. Inhismemoirs of 1903-1916 Levi-Civita contributed to celestialmechanics inthestudy of the three-body problem: the determination of the motionof three bodies, considered as reduced to their centers of mass andsubject to mutual Newtonian attraction. In 1914-1916 he succeededin eliminating thesingularities present atthepointsof possiblecollisions.past or future. His research in relativity led Levi-Civita to mathematical problems suggested by atomicphysics, whichin the 1920s was developing outsidethe traditional framework: the general theory of adiabatic invariants, themotionof a bodyof variable mass,the extension of theMaxwellian distribution to a systemof corpuscles, andthe SchrOdinger equation.

The covariant differentiation of vectors can be extended to arbitrary tensors once it is defined for l-forms. For this, we make the extra assumption that Vu can be "pushed" inside a pairing, and that when acting on a tensor product, it obeys the product rule of differentiation, i.e., Vu acts as a derivation on the algebra of tensors. Let w be a I-form and v a vector. Then Vu (w, v) = (Vuw, v)

+ (w, Vuv)

(28.34)

defines the covariant derivative of w. Using Equation (28.34) the reader may readily show the primary covariant derivative relation: (28.35) Since an arbitrary tensor of a given kind can be expressed as a linear combination of tensor product of vectors and l-forms, our knowledge of the action of Vu on functions (coefficients of expansion), vectors, and l-forms, plus the assumed derivation property of Vu, is enough to uniquely define the action of Vu on any tensor. Equation (28.34) shows that the covariant derivative of a l-form with respect

gOO

28. DIFFERENTIAL GEOMETRY

to a vector is another I-form, because when it pairs up with a vector it gives a number. We have already pointed out that the covariant derivative of a vector with respect to another vector is a third vector. We therefore conclude that

28.3.3. Box. The covariant derivative ofa tensor ofa given kind is another tensor of the same kind.

28.3.4. Example. Using the definition of the covariant derivative, the reader may check that Note the change in order of indices! Now consider two bases

(28.36)

Ier} and {eil}. Write the primed basis in terms of the other: ei' =

RJ;,ej. Then ~

ek,r i']'

== Vej'ei' =

.

I

.

VRI.,el (RJi,ej) = R pVel (RJj,ej) J

=Rip (Rj'f V" =

(ej)

+ V" (R~f) ej I

Rlj'R~,emrjl + R1l' el (Ri,) em. '-,-'

=R~"l

Connection coefficients are not tensors!

=

Writing ek' Ric/em on the LHS, equating the components on both sides, and multiplying both sides by the inverse of the transformation matrix R, we obtain

r '"i' i'

I

j

kf

m

m

1

Rkfm R i' R i' r jl + R m R j' R 1',1' =-,,how a (1, 2)-tensor transforms

(28.37)

nontensorial term

where Rk~ ea (R-1)k1m' Equation (28.37) shows that the connection coefficients aTe not II tensors.

28.3.5. Example. Equation (28.36) connects r ij k with the structure constants of the Lie algebra of vector fields on a manifold. To see this connection, use the first property of the covariant derivative in Theorem 28.3.2 and Equation (28.36) to obtain

lei. ejl = Veiej - Vejei =

ek

(r kj i

-

r~j).

It follows from this equation that

C~j

=

rkj i -

r~j'

(28.38)

In particular, in a coordinate frame, C~j = 0 and the connection coefficients are symmetric in their lower indices, a result we obtained earlier. II

28.3 COVARIANT DERIVATIVE AND GEODESICS

gradient operator for

901

Let T be a tensor of type (r, s). It is convenientto think of Vu T as the contraction of another tensor S '" VT with n. Then V may naturally be treated as a linear operatorthat maps TJ (M), the bundleoftensorfields oftype (r, s), to TJ+! (M), the bundle of tensor fields of type (r, s + I). One then writes V : TJ (M) ---> T;+l (M) and calls V the generalized gradient operator. If

tensors

then (28.39)

and, withu =

ukek,

VuT = T~l ..' : ukej ® ... @ej ® eit ® ... ® e is . ll ... ]s,k 1 r

(28.40)

Using these relations, we can calculate the components of the covariant derivative of a general teosor. It is clear that if we use ek instead of u, we obtain the kth

component of the covariant derivative. So, on the one hand, we have (28.41) and on the other hand, V ek T = V ek (Til"'" e· <81 ••• <81 e·lr <81 €iI <81 ••• <81 €J,) It ... }s '1

= T~l 'r keil ® ... ® ej ® eit ® ... ® e is It ls, T

r

+ T;11......~: L m=l s

eit

® ... ®

Vekeim

® .. . CiT ® e it ® ... ® eis

~ enr

imk

+ T~l1l ls'r L....J '""" ej1 ® ... ®ejr ® €iI ® ... ® m=l

Vek ~.im ~

-rJ;;ken

... ® €is, (28.42)

by (2S.35)

where Til"'" k es ek (T il ...i, ) . Equating the components of Equations (28.41) and n-.». n-;» (28.42) yields

(28.43) where only the sum over the subindex m has been explicitly displayed; the (hidden) sum overrepeated indicesis, as always,understood.

902

28. DIFFERENTIAL GEOMETRY

There is a useful relation between the covariaut aud the Lie derivativethat we derivenow.First, letTbe of type (2, 0) audwriteitinsomeframeasT = Tijei®ej. Apply the covariautderivativewith respect to u to both sides to obtain VuT = u(Tij)ei ® ej derivation ofthe relation between the Lie and the covariant derivatives

+ Tij (Vuei) ® ej + Tijei ® (Vuej).

Similarly, LuT = n(Tij)ei ® ej

+ Tij (Luei) ® ej + Tijei e (Luej).

Now use Luej = [u, ej] = Vuej - Vejn to get

t.; T = v,T -

Tij [(Vein) ® ej

+ ei e

(28.44)

(Vejn)].

Ou the other haud, if we apply Vu aud L u to both sides of 8~ = (.i, ej) aud use [u, eil = Vuei - VeiU' we obtain Vue i = Lue i - (Vej u)' e j

.

I!follows that for T = T;j.i ® .j, we have LuT

. k' " k = VuT + T;j [(Ve,n)'. ®.J + (Ve,n)J .' ®. ].

(28.45)

One cau use Equations (28.44) aud (28.45) to generalize to a tensor of type (r, s). If n happens to be taugent to a curve t ~ y(t), Equation (28.40) is written as DT~·l ... ~r

VuT =

JI..·], e.I] ® ... ® e,'r ®.h ® ... ® .j, • dt

(28.46)

where DT~I".i! fdt '" Til".i, uk. In a coordinate frame, with u i = xi = dx i fdt, EquationsJC i'lt40) aud (qlf.fjj give DT~l ...~r

1l •.. Js

=

. .

]s;kdt

lI

dt -

dxk

'r _

T~I

,

"

dT~l ...i!

=

}I···js

..

r .. . +~ T~1 ...I.III-1~lm.+l

L...J

dt

m=l

.

~ TII .•• lm-llmlm+] •.. lrrn

L...J

j!··.jm-In}m+I···}s

m=l

im k

dx k _

dt

. ...lr .

Jl···JIII-I1mJm+l···J.~

. dxk

rim ,_ nk

dt

(2847) .

For the case of a vector (28.47) becomes Dv k dt

dv k dt

. k dx' n-r-r-', J' dt

--=-+vJr

(28.48)

This is au important equation, to which we shall return shortly. Withthe generalizedgradientoperatordefined,we caucoustructthe divergeuce of a tensor just as in vector aualysis. Given a vector, the divergeuce operator V· acts on it aud gives a scalar, or, in the lauguage of tensor aualysis, it lowers the upper indices by 1. This takes place by differentiatingcomponentsaud contracting the upper index with the newlyintroducedindex of differentiation. The divergence of au arbitrary tensor is defined in precisely the same way:

903

28.3 COVARIANT DERIVATIVE AND GEODESICS

divergence ofa tensor field

28.3.6. Definition. Given a tensor field T, define its divergence V . T to be the tensor obtained from VT by contracting the last upper index with the covariant derivative index. In components,

The covariant derivative replaces the ordinary derivative when functions are generalized to tensor fields, and in many respects it is very similar to ordinary derivatives. The aspect of the covariant derivative that is in contrast to the ordinary derivative, namely, its lack of commutativity, is related to an innportant geometrical object that we encountered before: curvature. Let us find this relation. Start with

an ordinary l-form w = (Vie;, and note that (dw, u A v) = (dwi A ei + Widei, u A v) = (dwi A ei, u A v) + (Widei, U A v) ,

(28.49) where u and v are arbitrary vectors. For the first term of Equation (28.49), we obtain (dwi A

e, U A v) = ido», u) vi -

ido», v) u i

= ViU(Wi) -

UiV(Wi)'

Now we use

U«(v, w)) = U(ViWi) = WiU(Vi) v«(u, w»

+ ViU(Wi),

= V(uiWi) = WiV(Ui) + UiV(Wi),

to get idto; A ei, U A v) = uf (v, w) -V«U, w»

+ Wi[V(Ui) -

u(v i)].

(28.50) With dei

= _wi j

A

ej

= -rijkek A ej, the second term of (28.49) becomes

(widei, U A v) = -Wirijk (ek A e j, U A v) = -Wirijk (ukv j - u j vk)

= -Wiukvj (r i j k - r ikj) = _ukv j (w, [ek, ej]). ~

(£i,[ek,ejD

We also note that [U, v]

= [ukek, viej] = ukek(vjej) - vjej(ukek) = [u(vj)]ej + ukvjekej - [v(uk)]ek - ukVjejek = [u(v j) - v(uj)]ej

+ ukvj[ek, ej].

904

28. DIFFERENTIAL GEOMETRY

It follows that

(28.51) aod the second term of (28.49) becomes (Widei, u A v) = - (w, [u, v])

+ Wj[u(v j) -

v(u j)].

(28.52)

Combining Equatious (28.49), (28.50), aod (28.52), we obtain (dw, u

A

v) = u«v, w)) - v«u, w)) - (w, [u, v]) .

(28.53)

Equation (28.53) is the basis of our generalization involving the covariant derivative: We replace w with a vector-valued l-form T, aod the derivatives u(.. · ) aod v(···) with Vu("') aod Vv("')' The result will be (dT, u A v) = Vu«v, T)) - Vv«u, T)) - (T, [u, v]).

(28.54)

Choose a vector w aod replace T with dw to obtain (d 2w, U A v) = Vu «v. dw)) -Vv «u, dw)) - (dw, [u, v]) '-v-'" ;;:=Vvw

"

="Vuw

=V[u,vjW

= vuvvw - vvvuw - Vlu,vlW es R(u, v)w,

(28.55) where we have iotroduced the linear operator R(u, v) : X(M) -> X(M) in the last Iine, We rewrite the definition of this operator as

lI(u, v) = VuVv - VvVu - Vlu,vl'

(28.56)

Equations (28.55) aod (28.56) cao be used to derive the following identity (see Problem 28.11): R(u, v)w

+ R(w, u)v + R(v, w)u =

O.

(28.57)

Define the map R : X* (M) x X(M) x X(M) x X(M) -> R by R(w, w, u, v) sa (w, R(u, v)w) . Riemann curvature "tensor'ts indeed a tensor inthe sense of Box 26.4.11 !

(28.58)

The covariaot derivatives in the definition of R(u, v) may give the impression that R will differentiate whatever appears to its right. It is a remarkable property of R that this will not happen: That R is a tensor of type (1,3) [see Box 26.4.11], the Riemann curvature tensor, follows from Equations (28.7) aod (28.55). The reader may verify that the components of this tensor are precisely those iotroduced in Equation (28.22). In fact, it is now more appropriate to change Oii to Rij aod write (28.59)

905

28.3 COVARIANT OERIVATIVE AND GEODESICS

From Equation (28.57) follows the cyclic property of the lower indices of the components of R as given in Equation (28.24). The components of R can be conveniently evaluated in terms of connection coefficients: Ri j kl ea R(.i, ej, ek, el) = (.1, R(ek, e/)ej) i

= (e , VekVe1ej - Ve/Vekej - V[ebel1 e j )

=

(28.60)

<.1, V' kV./ej) - <.i, V./ V.kej) - <.i, V['ko./lej) .

We calculate each pairing separately. The vector in the first paiting is V'kV./ej = V' k (rjle,.) = [ek (rjl)]e,. =

+ r"jlv.ke,.

rjt,kern + rjlrnmken.

Similarly, the second vector can be expressed as Vel Vekej =

rjk,lem + rjkrnmlen.

For the third vector, we use the definition of the structure constants and write

Substituting these in Equation (28.60), we obtain i R j k1

=

fi j 1. k - fijk,l

+ rjtrimk - fjkriml - C11 r ij m '

(28.61)

If we use coordinate frames, the structure constants arezero, and we get . R'jkl

ar i j l ar i j k · . = - k - - --1- + r1mkfjl - r1m1fjk'

ax

ax

(28.62)

If the manifold has a metric (Riemannian or pseudo-Riemannian manifold), then a combination of (28.26) and (28.62) gives the curvature in terms of the mettic

tensor.

28.3.2 Geodesics The reader may recall from elementary physics courses that in a flat space, one is allowed? to move a vector about as long as it is kept parallel to itself. In any kind of parallel displacement of a vector, one moves the (tail of the) vector along some curve. This curve is in our subconscious in flat space, because it plays no role in such a displacement-all curves give the same end result. However, in curved spaces, different curves "parallel transport" a vector differently, so that the "Except in situations wherethepositionof thevectoris important, as in torques andangular momenta.

906

28. DIFFERENTIAL GEOMETRY

end result will be different. Each curve honestly "thinks" that it is indeed parallel transporting the vector, i.e., that it is not changing the direction of the vector as

thelatter moves alongtheformer. No change meanszeroderivative; and sincethe concept of derivative is local, let us replace the curve with its Iangent u: parallel transport of vectors and tensors along a vector

28.3.7. Definition. A vector v is said to be paralleltransportedalong u ifVuv = O. Similarly, a tensor T is said to be parallel transported along u ifVuT = O.

If the manifold M has a metric g, then it is desirable for the parallel transportation not to affect the metric, i.e., that the angle between any two vectors remain the same after transportation. Therefore, we demand that 9 be constant for parallel transportation along any vector, i.e., Vug

=0

\I u

=}

Vg

= 0,

or

g;j;k

=0

\I i, j, k.

(28.63)

A statement equivalent to this equation is that

28.3.8. Box. The operation of raising and lowering of indices commutes with the operation of covariant differentiation.

Consider two vectors v and w. If 9 satisfies Equation (28.63), then Vu(v' w) = Vu[g(v, w)] ea Vu (g, v <81 w) = (Vug, v <81 w) + (g, (Vuv) <81 w)

= g(Vuv, w)

+ (g, v <81 (Vuw))

+ g(v, Vuw),

or Vu(v· w) = (Vuv) . w + v (Vuw).

(28.64)

In particular, if v and w are parallel transported along u, their dot product will not change. In the flat space of a large sheet of paper, construction of a straight line in a given direction starting at a given point Po is done by laying down the end of a vector (a straight edge) at Po pointing in the given direction, connecting Po to a neighboring point PI aloog the vector, moving the vector parallel to itself to PI, connecting PI to a neighboring point Pz, and continuing the process. In the language of the machinery of the covariant derivative, we might say that a straight line is constructed by transporting the tangent vector parallel to itself.

geodesics defined

geodesic equation

28.3.9. Definition. Let M be a manifold and y : [a, b] --* M a curve. Then y is called a geodesic of M if the tangent vector u at every point of y is parallel transported along itself: Vuu = O. It follows from Equation (28.48}--with v k

= uk = dx k / dt-that a geodesic

907

28.3 COVARIANT DERIVATIVE AND GEODESICS

curve satisfies the following DE: d 2xk dx i dx j -+~.--=O dt 2 J' dt dt .

~~

This second-order DE, called the geodesic equation, will have a unique solution if xi (a) and xiCal, i.e., the initial point and the initial direction, are given. Thus,

28.3.10. Box. Through a given point and in a given direction passes only one geodesic curve.

28.3.11. Example. Let us determine the geodesics of the space whose arc length is given hyds 2 = (dx 2 +dy2)/y2 (see Example 28.2.5). With x = x t andy =x 2, we recognize themetric tensor as 1

su = g22 = 2' Y gtl = g22 = y'2

g12 = g21 = 0,

Using Equation (28.26), we canreadily calculate the nonzero connection coefficients: I

J'nz

= f12l = -f2tl = f222 = -3"' y

The geodesic equation for the first coordinate is 2x i dx i d_+f1. dx _ _ =0 2 dt J' dt dt '

or

e. 1 (dx)2 1 dx dy 1 (d y)2 dt 2 + I' tl dt + 2f 12 dt dt + I' 22 dt . = O. Tofindthe connection coefficients withraised indices,we multiply thosewith all indices lowered by theinverse of themetric tensor. Forinstance, f\2

= gli f i12 = gtl f

tl2

+

gI2 f212 ~

= y2 (-~) = y

_.!:.. y

~O

Similarly, fltl

= 0 = fI22 , and the geodesic equation for the first coordinate hecomes

d 2x 1 dx dy - - 2 - - - =0. dt 2

y dt dt

(28.66)

Forthesecondcoordinate, we need r 2I1•r 212•andr 222·Thesecanbereadilyevaluated as above, withtheresult 2 2 1 f tl =-f 22 = - , Y

908

28. DIFFERENTIAL GEDMETRY

yielding the geodesic equation for the second coordinate 2y d dt 2

+ .!. (dx)2 _.!. (d y)2 = o.

With

Y dt

(28.67)

Y dt

x '" dxldt, Equation (28.66) can be writtenas

dx _ 2Xdy/dt = 0 dt y

'* dx/dt = 2 dy/dt '* x = ci. x

Using the chain rule and the notation y'

Y

== dyjdx, we obtain

dy t , C 2 / dt = Y x = Y y, d 2Y d dy' -2 = C(2y..2'.y' + i - x ) = C 2(2y3yl2 + y4 y"). dt dt dx

Substituting in Equation (28.67) yields y3y'2

+ y4y" + y3 =

0

'* (y')2 + yy" + I = 0 '* ..'£(yy') + 1= O. dx

It follows that ss' = -x + A and x 2 + y2 = 2Ax arbitrary radii whose centers lie on the x-axis.

+ B. Thus, the geodesics are circles with II

28.4 Isometrles and Killing Vector Fields Whenever a structure is defined on a given family of sets, mappings between snch sets that preserve that structure become the "natural" mappings. Thns, the natural maps among vector spaces are linear maps, among groups are homomorphisms, among algebras are algebra homomorphisms, and among manifolds are smooth maps. The introdnction of a metric on a manifold makes certain maps more privileged than others. isometry defined

Isometries of a manifold form a subgroup of Diff(M).

Killing vector field

28.4.1. Definition. Let M and N be Riemannian manifolds with metrics 9M and 9N, respectively. The smooth map 1/1 : M -> N is called isometric at P E M if 9M(X, Y) = 9N(1/I.X, 1/IS)for all X, Y E Tp(M). An isometry of M to N is a bijective smooth map that is isometric at every point of M, in which case we have 1/I.9M = 9N' Of special interest are isometries 1/1 : M -e- M of a single manifold. These happen to be a subgroup of Diff(M), the group of diffeomorphisms of M. 10 the subgroup of isometries, we concentrate on the one-parameter groups of transformations. These define (and are defined by) certain vector fields: 28.4.2. Definition. Let X E X(M) be a vector field with integral curve F,. Then X is called a Killing vectorfield if F, is an isometry of M.

909

28.4 ISOMETRIES ANO KILLING VECTOR FIELDS

The following proposition follows innnediately from the definition of the Lie derivative [Eqnation (26.29)]. 28.4.3. Proposition. A vector field X E X(M) is a Killing vectorfield O.

if and only

if Lxg =

Choosing a coordinate system {xi}, we write 9 = gijdx i <81 dx! and conclude that Xl j is a Killing vector field if and only if

a

0= Lx(gijdx i <81 dx j) = X(gij)dx i <81 dx! Killing equation

+ gij(LXdxi) <81 dx j + gijdx i <81 (Lxdxj).

Using Equation (26.33) for the I-form dx k, we obtain the Killing equation (28.68) If in Equation (28.45) we replace T with 9 and u with X, where X is a Killing vector field, we obtain

(28.69)

Killing equation in terms ofcovariant derivative

where we have assumed that the covariant derivative is compatible with the metric tensor. The reader may check that Equation (28.69) leads to Xj;k

+ Xk;j

(28.70)

= O.

This is another form of Killing equation. 28.4.4. Proposition. IIX is a Killing vector field, then its inner product with the tangentto any geodesic is constantalong that geodesic, i.e., ifu is such a tangent, then Vu[g(u, X)] = O. Proof We write the desired covariant derivative in component form, k i-o i k t-o i u (gijU Xl);k = U gij;k U Xl

" + Uk gij Ui;k.Xl + Uk su» jXi;k

'-,-'

'-,-'

=0 1

= 2: (Xi;k

=0

i k

+ Xk;i) U U

,

'-,---'

=0 by (28.70)

where the first term vanishes by assumption of the compatibility of the metric and the covariant derivative, and the second term by the geodesic equation. D 28.4.5. Example. In a flat m-dimensional manifold we canchooseanorthonormal coordinateframe (Theorem28.2.9), so that the Killingequationbecomes

(28.71)

910

28. DIFFERENTIAL GEOMETRY

Setting i = j, we see that OjXj = 0 (no sum). Differentiating Equatioo (28.71) with respect-to xi , we obtain

-

2 0iXj+OjOiXi =0

=}

2 0iXj =0

V

t, j.

~O

Therefore, X j is linear in xi, i.e., Xj = a jkXk + bj with a jk and bj arbitrary. Inserting this in (28.71), we get aij + aji = O. The Killing vector is thee

x = (aijx i + hi) aj = maximally symmetric spaces

!a i j (xi aj

-

Xiaj)

+ biaj.

The first term is clearlythe generator of a rotation andthe secondterm that of translation. Altogether, there are m(m - 1)/2 rotations and m translations. So the total number of independent Killingvectors is m(m + 1)/2. Manifolds thathavethis manyKillingvectors are called maximally symmetric spaces. IIlII Using Equations (26.34) and (26.35) one can show that the set of Killing vector fields Ona manifold form a vector subspace ofX(M), and that if X and Y are Killing vector fields, then so is [X, V]. Thus,

28.4.6. Box. The set ofKilling vector fields forms a Lie algebra.

28.4.7. Example. From 9 = dB (i!)dB +sin2 Bdcp (i!)dcp, the metric of the unit sphere S2, one writesthe Killing equations iJeXe o~X~

+

O~X.

oeX~

+

+ iJeXe =

0,

(28.72)

+ 2sinBcos BXe = 0,

o~Xe - 2cotBX~ =

Thefirst equation impliesthatX() equation yields

X~ = -iF(cp) sin2B + g(B),

O.

= !(rp),

a function of ip only.Substitution in thesecond

where

dF f(cp) = dcp'

Inserting this in the third equation of (28.72), we obtain dg df 1 -F(cp)cos2B+ dB + dcp +2cotB[ZF(cp)sin2B-g(B)] =0, or

. [d g dB - 2cotBg(B)]

df

+ [dCP + F(CP)] = O.

Forthisequation to holdfor all () andtp, we musthave dg df dB - 2 cot Bg(B) = C = - dcp - F(cp),

28.4 ISOMETRIES AND KILLING VECTOR FIELDS

911

where C is a constant. Thisgives g(O)

= (Cl

- C cot 0) sin z 0

and

/('1') = Xe = A sin 'I'

+ B cos 'I'

with

x'" = (A cos 'I' -

B sin '1') sinO cose

+ Ct

sin z O.

A general Killing vector fieldis thus givenby

where

= -coscpoe + cot 61 sinrporp.

Lx

L y =sinqJoe +cotBcoscp(}cp, Lz =

a",

.

are the generators of SO(3).

discussion of conformal transformations and conformal Killing vector fields

Sometimes it is useful to relax the complete invariance of the metric tensor under the diffeomorphism of a manifold induced by a vector field and allow a change of scale in the metric. More precisely, we consider vector fields X whose flow F , changes the metric of M. So F,.g = e¢(I)g, where e is a a real-valued function on M that is also dependent on the parameter t. Such a transformation keeps angles unchanged but rescales all lengths. In analogy with those of the complex plane with the same property, we call such transformations conformal transformations. A vector field that generates a conformal transformation will satisfy

k

X aUij

· + aiX kgkj + ajXk gki = -1/Jgij,

a"'j 1/J='at

1=0

,

(28.73)

and is called a conformal Killing vector field. Wenow specialize to a flat m-dimensional manifold and choose an orthonormal coordinate frame (Theorem 28.2.9). Then Equation (28.73) becomes

aiXj Remember Einstein's summation convention!

+ ajXi

= -1/Jgij.

(28.74)

Multiply both sides by gij and sum over i to obtain

.

za' Xi = -m1/J

=}

. a' Xi

m

= -21/J,

m=dimM.

Apply ai to both sides of Equation (28.74) and sum over i. This yields (28.75)

912

28. DIFFERENTIAL GEDMETRY

Differentiate both sides of the second equation in (28.75) with respect to x k and symmetrize the result in j and k to obtain (m - 2)aiak1ft

= aiai(akXi + aiXk) = -gikaiai1ft.

Raising the index j and contracting it with k gives ai ad = 0 if m that

t= I. It follows (28.76)

t=

Equations (28.74), (28.75), and (28.76) detemnine both 1ft and Xi if m 2. It follows from Equation (28.76) that 1ft is linear in x, and, consequently [from (28.74)], that Xi is at most quadratic in x. The most general solution to Equation (28.74) satisfying (28.75) and (28.76) is (28.77)

conformal group

dilitation inversion, or special conformal transformation

where aij = -a ji andindices of the constants areraised and loweredas usual. Equation (28.77) gives the generators ofan [(m+ 1)(m+2)/2]-parametergroup of transformations on Rm , m 2 called the conformal group. The reader should note that translation (represented by the parameters bi ) and rotations (represented by the parameters aij) are included in this group. The other finite (as opposed to infinitesimal) transfomnations of coordinates can be obtained by using Equation (27.25). For example, the finite transfomnation generated by the parameter 8 is given by the solution to the DE dx'! [dt = x'l , which is x'i = e'x i, or x'i = e"xi, and is called a dilitation of coordinates. Similarly, the finite transfomnation generated by the parameter ci is given by the solution to the DE dx'i [de, = 8ij x'kx' - 2X'i x' j or

t=

k

'

which is called inversion, or the special conformal transformation. Equations (28.75), and (28.76) place no resttiction on 1ft (and therefore on Xi) when m = 2. This means that .

28.4.8. Box. The conformal group is infinite-dimensional for R 2 .

In fact, we encountered the confomnal transfomnations of R 2 in the context of complex analysis, where we showed that any (therefore, infinitely many) analytic function is a confomnal traosfomnation of C ~ R 2 • The confomnal group of R 2 has important applications in string theory and statistical mechanics, but we shall not pursue them here.

28.5 GEODESIC DEVIATION AND CURVATURE

913

28.5 Geodesic Deviation and Curvature Geodesics are the straight lines of general manifolds on which, for example, free particles move. If u represents the tangent to a given geodesic, one can say that V.u = 0 is the equation of motion of a free particle. In flat spaces, the relative velocity of any pair of free particles will not change, so that their relative acceleration is always zero. In general, however, due to the effects of curvature, we expect a nonzero acceleration. Let us elaborate on this. Consider some. region of the manifold through whose points geodesics can be drawn iu vatious directions. Concentrate on one geodesic and its neighboring geodesics. Let t desiguate the parameter that locates points of the geodesic. Let s be a continuous parameter that labels different geodesics (see Fignre 28.1). One can connect the points on some neighboring geodesics corresponding to the same value of t and obtain a curve paramettized by s. The collection of all geodesics that pass through all points of this curve form a two-dimensional submanifold with coordinates t and s. Each such geodesic is thus described by the geodesic equation V.u = 0 with u = a/at (because t is a coordinate). Furthermore, as we hop from one geodesic to its neighbor, the geodesic eqnation does not change; i.e., the geodesic eqnation is independent of s. Translated into the langnage of calculus, this means that differentiation of the geodesic equation with respect to s will give zero. Translated into the higher-level language of tensor analysis, it means that covatiant differentiation of the geodesic equation will yield zero. We write this final translation as V.(V.u)= 0 where n es a/as. This can also be written as 0= V. V.u = V. V.u + [V., V.lu = V.(V.n + [n, uJ)

+ [V., V.Ju,

where we have used the first property of Theorem 28.3.2. Using the fact that n and u are coordinate frames, we conclude that [n, u] = 0 which in conjnnction with Equation (28.56) yields V. V.n + R(u, u)u = O. relative acceleration

equation ofgeodesic deviation

(28.78)

The first term can be interpreted as the relative acceleration oftwo geodesic curves (or free particles), because V. is the generalization of the derivative with respect to t, and V.n is interpreted as relative velocity. In a flat manifold, the relative acceleration for any pair of free particles is zero. When curvature is present, it produces a nonzero relative acceleration. By writing u = u' a, and n = n kak and substituting in Equation (28.78), we arrive at the equation of geodesic deviation in coordinate form:

. . k

d 2nm

u1uJn Rijk = - 2 -

dt

d

+-

dt

. k

(ut n

j

dn

k

. . k

r"kj) + U df'kj +u1uJn t

rnkjr~i' (28.79)

where we have used the fact that u'a'f = df/dt for any function defined on the manifold.

914

28. DIFFERENTIAL GEOMETRY

s

u t

Figure 28.1

A region of the manifold and the two-dimensional surface defined by s and

t.

Einstein's interpretation of gravity

The chain that connects relative acceleration to curvatnre has another link that connects the latter two to gravity. From a Newtonian standpoint, gravity is the only force that accelerates all objects at the same rate (equivalence principle). From a geometric standpoint, this property allows one to include gravity in the structnre of space-time: An object in free fall is considered "free," and its path, a geodesic. Locally, this is in fact a better picture of reality, because inside a laboratory in free fall (such as a space shuttle in orbit around earth) one actually verifies the first law of motion on all objects floating in midair. One need not include an extennal phenomenon called gravity. Gravity becomes part of the fabric of spacetime. But how does gravity manifest itself? Is there any observable effect that can indicate the presence of gravity, or by a mere transfer to a freely falling frame have we been able to completely eliminate gravity? The second alternative wonld be strange indeed, because the source of gravity is matter, and if we eliminate gravity completely, we have to eliminate matter as well! If the gravitational field were homogeneous, one could eliminate it-and the matter that produces it as well, but no such gravitational field exists. The inhomogeneity of gravitational fields has indeed an observable effect. Consider two test particles that are falling freely toward the source of gravity on two different radii. As they get closer and closer to the center, their relative distance-in fact, their relative velocity--changes: They experience a relative acceleration. Since as we saw in Equation (28.78), relative acceleration is related to curvature, we conclude that

28.5.1. Box. Gravity manifests itself by giving space-time a curvature.

28.5 GEODESIC DEVIATION AND CURVATURE

915

This is Einstein's interpretation of gravity. From a Newtonian standpoint, the relative acceleration is caused by the inhomogeneity of the gravitational field. Such inhomogeneity (in the field of the Moon and the Sun) is responsible for tidal waves. That is why the curvature term in Equation (28.78) is also called the tide-producing gravitational force.

28.5.1 Riemann Normal Coordinates

Riemann normal coordinates

Starting with a point P of an n-dimensional manifold M on which a covariant derivative is defined, we can construct a unique geodesic in everydirection, i.e., for every vector in 'Jp(M). By parallel transportation of the tangent vectors at P, we can construct a vector field in a neighborhood of P: The value of the vector field at Q-assumed to be close enough to P-is the tangent at Q on the geodesic starting at P and passing through Q. 8 The vector field so obtained makes it possible to define an exponential map from the tangent space to the manifold. In fact, the integral curve exp(tX) of any tangent vector X in 'J p (M) is simply the geodesic associated with the vector. The uniqueness of the geodesics establishes a bijection (in fact, a diffeomorphism) between a neighborhood of the origin of 'J p(M) and a neighborhood of P in M. This diffeomorphism can be used to assign coordinates to all points in the vicinity of P. Recall that a coordinate is a smooth bijection from M to JR". Now choose a basis for 'J p (M) and associate the components of tX in this basis to the points on the geodesic exp(tX). Specifically, if {ai}i~l are the components of X in the chosenbasis,then

i = 1,2, ... , n, are the so-called Riemann normal coordinates (RNes) of points on the geodesic of X. The geodesic equations in these coordinates become

r kjia i a!. = 0

=}

r kji

= 0

because r~i is symmetric in i and

i-

28.5.2. Proposition. The connection coefficients at a point P E M vanish in the

Riemann normalcoordinates at P. Using Equation (28.43), we immediately obtain the following: 28.5.3. Corollary. Let T be a tensor field on M with components TJ:......;; with respect to a Riemann normal coordinate system {xi} at P. Then

..

T~l ...':

ll···]s;k

=

a · ·'r.

_T~l ...

8xk n.:»

SWe are assuming that through any two neighboring points one can always draw a geodesic. For a proof see [Koba 63,

pp. 172-175].

916

28. DIFFERENTIAL GEOMETRY

Riemann normal coordinates are very useful in establishing tensor equations. This is because 28.5.4. Box. Two tensors are identical the same in any coordinate frame.

if and only if their components are

Therefore, to show that two tensors fields are equal, we pick an arbitrary point in M, erect a set of RNCs, and show that the components of the tensors are equal. Since the connection coefficients vanish in an RNC system,andcovariant derivatives are the same as ordinary derivatives, tensor manipulations can be simplified considerably. For example, the components of the curvature tensor in RNCs are i i _ jl R jkl - - k - -

ar

ax

ar i j k

(28.80)

-1-'

ax

This is not a tensorrelation-the RHS is not a tensor in a general coordinate system. However, ifwe establish arelation involving the components of the curvature tensor alone, then that relation will hold in all coordinates, i.e., it is a tensor relation. For instance, from the equation above one inunediately obtains i R j kl

+ R i1j k + R ik1j

= O.

Since this involves only a tensor, it must hold in all coordinate frames. This is the Bianchi identity of Equation (28.24). 28.5.5. Example. Differentiate Eqnation (28.62) with respect to x m and evaluate the result in RNC to get Rijkl;m

second Bianchi identity

= Rijkl,m = rijl,km -

rijk,lm'

From this relation and r iJ, "I k m = ri"1 k' we obtainthe second Bianchi identity: J ,m Rijkl;m

+ Rijmk;1 + Rij1m;k =

0

and

Rijlkl;m] =

O.

(28.81)

In Einstein's general relativity, this identity is the analogue of Maxwell'spairof homogeneons equations; F.p,y + Fy.,p + Fpy,. = O. III

28.5.2 Newtonian Gravity The equivalence principle, relating gravity with the curvature of space-time, is not unique to Einstein. What is unique to him is combining that principle with the assumption of local Lorentz geometry, i.e., local validity of special relativity. Carlan also used the equivalence principle to reformulate Newtouian gravity in the language of geometry. Rewrite Newton's second law of motion as F=ma

=}

a= -V

=}

a2 x j

a

--+--. at 2 ax} =0,

28.5 GEODESIC DEVIATION AND CURVATURE

917

where 4> is the gravitational potential (potential energy per unit mass). The Newtonian universal time is a parameter that has two degrees of freedom: Its origin and its unit of measurement are arbitrary. Thus, one can change t to t = ar + b without changing the physics of gravity. Taking this freedom into account, one can write

(28.82) Comparing this with the geodesic equation, we can read off the nonzero connection coefficients:

r iOO =

a4>" ax J

j = 1,2,3.

(28.83)

Inserting these in Equation (28.62), we find the nonzero components of Riemann

curvature tensor: , . a24> R JOkO -- - R JOOk -- axiaxk'

(28.84)

Contraction of the two nonzero indices leads to the Laplacian of gravitational potential

Therefore, the Poisson equation for gravitational potential can be written in terms of the curvature tensor: Roo sa R ioi O = 4rrGp,

Ricci tensor defined

(28.85)

where we have introduced the Ricci tensor, defined as Rik sa

R~jk'

(28.86)

Equations (28.83), (28.84), and (28.85) plus the law of geodesic motion describe the full content of Newtonian gravitational theory in the geometric language of tensors.? It is instructive to discover the relation between curvature and gravity directly from the equation of geodesic deviation as applied to Newtonian gravity. The geodesic equation is the equation of motion:

9Theclassic andcomprehensive book Gravitation, by Misner, Thorne, andWheeler, has a thorough discussionof Newtonian gravity in thelanguage of geometry in Chapter 12 and is highly recommended.

918

28. DIFFERENTIAL GEOMETRY

Differentiate this equation with respect to the parameter

s, noting

that

aIas

nialax i :

Now note that

ax j las = niax j laxi = nl . So, we obtain

a2n j

. a2 <1> at2 + n' axiaxj

= O.

This is equivalent to Equation (28.78), and one recognizes the second term as the tide-producing (or the curvature) term.

28.6

General Theory of Relativity

No treatment of Riemannian geometry is complete without a discussion of the general theory of relativity. That is why we shall devote this last section of the current chapter to a brief exposition of this theory. We have seen that Newtonian gravity can be translated into the language of differential geometry by identifying the gravitational tidal effects with the curvature of space-time. This straightforward interpretation of Newtonian gravity, in particular the retention of the Euclidean metric and the universality of time, leads to no new physical effect. Furthermore, it is inconsistent with the special theory of relativity, which mixes space and time coordinates via Lorentz transformations. Einstein's general theory of relativity (G1R) combines the eqnivalence principle (that freely falling objects move on geodesics) with the local validity of the special theory of relativity (that the metric of space-time reduces to the Lorentz-Minkowski metric of the special theory of relativity).

28.6.1 Einstein's Equation Since the metric plays a central role in general relativity, let us study its effect on the curvature tensor. The reader may verify that (Problem 28.23) 0= R(u, v)(w· x) = [R(u, v)w] . x + w· [R(u, v)x]. Using this, one can show that Rijkl

=-

(28.87)

Rjikl'

which is the first equation of (28.23). Equations (28.23) and (28.24) form a complete set of symmetries of the Riemarm curvature tensor. Other symmetries that follow from them are Rijkl

= Rklij

and

R[ijklJ

=

o.

(28.88)

28.6 GENERAL THEORY OF RELATIVITY

Einstein tensor

919

An important tensor that can be constructed out of Riemann curvature tensor and the metric tensor is the Einstein tensor G, which is related to the Ricci tensor defined in Equation (28.86). To derive the Einstein tensor, first note that the Ricci tensor is symmetric in its indices (see Problem 28.25): (28.89)

curvature scalar defined

Next, define the curvature scalar as i R -~ R i -~ gijR" IJ"

(28.90)

Now, contract i with m in Equation (28.81) and use the antisymmetry of the Riemarm tensor in its last two indices to obtain i R jkl;i

+ R jk;l -

R jl;k

= O.

Finally, contract j and I and use the antisymmetry of the Riemarm tensor in its first as well as its last two indices to get 2R ik;i - R;k 0, or Rjk;i - hjkR;i O. Summarizing the foregoing discussion, we write

=

V ·G=O,

=

(28.91)

Karl Schwarzschild (1873-1916) was the eldest of five sons andone daughter bornto Moses Martin Schwarzschild andhis wife, Henrietta Sabel. His father was a prosperous member of the business community in Frankfurt, with Jewish forbears in that city traced backto the sixteenth century. From his mother, a vivacious, warm person, Karl undoubtedly inheritedhis happy, ontgoingpersonality, and from his father, acapacity forsustained hard work. Hischildhood wasspent in comfortable circumstances among a largecircleof relatives, whose interests included art andmusic; he was the first to become a scientist. Afterattending a Jewish primary school,Schwarzschild entered themunicipal gymnasiumin Frankfurt attheage of eleven.His curiosity about theheavens was first manifested then: He savedhis allowance andboughtlensesto makea telescope. Indulging thisinterest, his father introduced him to a friend, J. Epstein, a mathematician who hada private observatory. With Epstein's son (later professor of mathematics attheUniversity of Strasbourg), Schwarzschild learned to use a telescopeandstudied mathematics of a more advanced type than he was getting in school. His precocious mastery of celestialmechanics resulted in two papers on double starorbits, written whenhe was barelysixteen. In 1891 Schwarzschild begantwoyears of studyat the University ofStrasbonrg,where Ernst Becker, director of the observatory, guidedthe development of his skills in practical astronomy-skills thatlater wereto form a solidunderpinning forhis masterful mathematical abilities.

920

28. DIFFERENTIAL GEOMETRY

At age twenty Schwarzschild went to the University of Munich. Three years later, in 1896, he obtained his Ph.D., summa cum laude. His dissertation was an application of Poincare's theory of stable configurations in rotating bodies to,several astronomical problems, including tidaldeformationin satellites andthe validity of Laplace's suggestion as to how the solar system had originated. Before graduating, Schwarzschild also found time to do somepractical workwith Michelson's interferometer.

At a meeting of the German AstronomicalSociety in Heidelbergin 1900he discussed thepossibilitythatspacewas non-Euclidean. In the sameyearhe published a paper giving a lower limit for the radius of curvatureof space as 2500 light years. From 1901 until 1909 he was professor at Gottingen, where he collaborated with Klein, Hilbert, and Minkowski, puhlishingon electrodynamics and geometricaloptics. In 1906,he studied the transportof energy through a star hy radiation. From Gcttingen he wentto Potsdam, butin 1914he volunteered formilitary service. He servedin Belgium,France, andRussia. While in Russiahe wrote two papers on Einstein's relativity theory andone on Planck's quantum theory. The quantum theory paper explained that the Stark effect could be proved from the postulates of quantum theory. Schwarzschild's relativity papers give thefirst exactsolutionofEinstein 's general gravitational equations, givinganunderstanding of thegeometry of spaceneara pointmass.He alsomadethefirst studyof black holes, showingthatbodiesof sufficiently large masswould haveanescapevelocityexceedingthespeedof light andso couldnotbe seen.However, he contracted anillness while in Russiaanddied soon afterreturning home.

The automatic vanishing of the divergence of the symmetric Einstein tensor has an important consequence in the field equation of GTR. It is reminiscent ofa similar situation in electromagnetism, in which the vanishing ofthe divergence of the fields leads to the conservation of the electric charge, the source of electromagnetic flelds.l" Just as Maxwell's equations are a generalization of the static electricity of Coulomb to a dynartrical theory, Einstein's GTR is the generalization ofNewtotrian static gravity to a dynamical theory. As this generalization ought to agree with the successes of the Newtonian gravity, Equation (28.91) must agree with (28.85). The bold step taken by Einstein was to generalize this relation involving ouly a single component of the Ricci tensor to a full tensor equation. The natural tensor to be used as the source of gravitation is the stress energy tensor!' TiLV ea (p

+ p)uiLUV + pgiLV,

or

T = (p

+ p)u 181 u + P9,

where the source is treated as a fluid with density p, four-velocity u, and pressure p. So, Einstein suggested the equation G = KT as the generalization of Newton's WIt wasMaxwell'sdiscovery of the inconsistency of thepre-Maxwellian equations of electromagnetism withcharge conservation that prompted him to change not only thefourth equation (to makethe entireset of equations consistent with the charge conservation), butalso the course of human history. 11 In GTR, it is customary to use theconvention that Greek indicesrunfrom0 to 3, i.e., theyinclude bothspaceandtime,while Latin indices encompass only thespacecomponents.

28.6 GENERAL THEORY OF RELATIVITY

Einstein's equation of the general theory of relativity

cosmological constant

921

universal law of gravitation. Note that V . G = 0 automatically guarantees massenergy conservation as in Maxwell's theory of electromagnetism. Problem 28.27 calculates K to be 8". in units in which the universal gravitational constant and the speed of light are set equal to unity. We therefore have

G=8".T,

or

R - !Rg = 8".[(p

+ p)u ® u + pg].

(28.92)

This is Einstein's equation of the general theory of relativity. The Einstein tensor G is nearly the only symmetric second-rank tensor made out of the Riemann and metric tensors that is divergence free. The only other tensor with the same properties is G + Ag, where A is the so-called cosmological constant (see Problem 28.28). When in 1917, Einstein applied his GTR to the universe itself, he found that the universe ought to be expanding. Being a fum believer in Natore, he changed his equation to G + Ag = 8".T to suppress the unobserved prediction of the expansion of the universe. Later, when the expansion was observed by Hubble, Einstein referred to this mutilation of his GTR as "the biggest blunder of my life."

Aleksandr Aleksandrovich Friedmann (1888-1925) was bom into a musicalfamily-his father, Aleksandr Friedmann, being a composer andhis mother, Ludmila Vojacka, the daughter of theCzech composer Hynek Vojacek. In 1906 Friedmann graduated from the gymnasium with the gold medal and immediately enrolled in the mathematics section of the department of physics and mathematics of St. Petersburg University. Whilestillastudent, he wrote anumber of unpublished scientific papers, one of whichwasawarded a gold medalby the department. Aftergraduation from theuniversity in 1910, Friedmann was retained in the department to prepare for the teachingprofession. In the fall of 1914, Friedmann volunteered for service in an aviation detaclunent, in whichhe worked, first on thenorthern front andlateron other fronts, to organize aerologic andaeronavigational services. While at thefront, Friedmann oftenparticipated in military flights as an aircraft observer. In the summer of 1917 he was appointed a section chief in Russia's first factory forthemanufacture of measuring instruments usedin aviation;' he later becamedirector of thefactory. Friedmann hadto relinquish this post because of the onset of heart disease. From 1918 until 1920, he was professor in the department of theoretical mechanics of Perm University. In 1920 he returned to Petrograd and worked at the mainphysics observatory of the Academy of Sciences,first as headof themathematics department andlater, shortly before his death, as director of the observatory. Friedmann's scientific activity was concentrated in the areas of theoretical meteorology andhydromechanics, where he demonstrated his mathematical talent andhisunwavering strife for, andabilityto attain, theconcrete, practical application of solutions to theoretical problems. Friedmann made a valuable contribution to Einstein's general theory of relativity. As always, his interest was not limited simply to familiarizing himself with this new field

922

28. DIFFERENTIAL GEOMETRY

of science butled to his own remarkable investigations. Friedmann's work on the theory of relativity dealtwith the cosmologicalproblem. In his paper "tiberdie Kriimmung des Raumes" (1922), he outlined the fundamental ideas of his cosmology: the supposition concerning the homogeneity of the distribution of matter in space and the consequent homogeneity and isotropy of space-time. This theory is especially important because it leads to a sufficiently correct explanation of the fundamental phenomenon known as the "red shift." Einstein himself thought that the cosmological solution to the equations of a fieldhadtobe static andhadto leadto a closedmodelof theuniverse. Friedmann discarded bothconditions andanived at an independent solution. Friedmann's interest in the theory of relativity was by no means a passingfancy. In the last years of his life, together with V. K. Frederiks, he beganworkon a multivolume text on modem physics. The first book, The World as Space and Time, is devoted to the theory of relativity, knowledge of whichFriedmann considered one of the cornerstones of aneducation in physics. In addition to his scientific work, Friedmann taught courses in highermathematics and theoretical mechanics at various colleges in Petrograd. He found time to create new and original courses, brilliant in their form andexceedingly varied in theircontent. Friedmann's unique coursein theoretical mechanics combined mathematical precision andlogical continuity withoriginal procedural andphysicaltrends. Friedmann diedoftyphoidfeverat theageofthirty-seven. In 1931,hewasposthumously awarded theLeninPrizeforhis outstanding scientific work.

28.6.2 Static Spherically Symmetric Solutions The general theory of relativity as given in Equation (28.92) has been strikingly successful in predicting the spacetime-? structure of our universe. It predicts the expansion of the universe, and by time-reversed extrapolation, the big bang cosmology; it predicts the existence of black holes and other final products of stellar collapse; and on a less grandiose scale, it explains the small precession of Mercury, the bending of light in the gravitational field of the Sun, and the gravitational redshift. We shall not discuss the solution of Einstein's equation in any detail. However, due to its simplicity and its use of geometric arguments, we shall consider the solution to Einstein's equation exterior to a static spherically symmetric distribution of mass. Let us first translate the two adjectives used in the last sentence into a geometric language. Take static first. We call a phenomenon "static" if at different instants it "looks the same." ThUS, a static solution of Einstein's equation is a spacetime manifold that "looks the same" for all time. In the language ofgeometry "looks the same" means isometric, because metric is the essence of the geometry of spacetime. In Euclidean physics, time can be thought ofas an axis at each point (moment) 12The reader maybe surprised to see thetwowords "space" and"time" juxtaposed withno hyphen; butthisis commonpractice in relativity.

28.6 GENERAL THEORY OF RELATIVITY

923

of which one cam assign a three-dimensional space corresponding to the "spatial universe" at that moment. In the general theory of retativity, space amd time cam be mixed, but the character of time as a parameter remains unaltered. Therefore, instead of am axis-s-a straight line-we pick a curve, a parametric map from the real tine to the mamifoldof space-time. This curve must be timelike, so that locally, when curvature is ignored amd special relativity becomes a good approximation, we do not violate causality. The curve must also have the property that at each point of it, the space-time manifold has the same metric. Moreover, we need to demamd that at each point of this curve, the spatial part of the space-time is orthogonal to thecurve.

stationary and static spacetimes; time translation isometries

Kiliing parameter

28.6.1. Definition. A spacetime is stationary if there exists a one-parameter group ofisometries F" called time translation isometries, whose Killing vector fields are timelikefor all t: g(e, > O. [fin addition, there exists a spacelike hypersurface E that is orthogonal to orbits (curves) of the isometries, we say that the . spacetime is static.

e

e)

We camsimplify the solution to Einstein's equation by invoking the syunnetry of spacetime discussed above in our choice of coordinates. Let P be a point of the spacetime manifold located in aneighborhood of some spacelike hypersurface E as shown in Figure 28.2. Through P passes a single orbit of the isometry, which starts at a point Q of E. Let t, the so-called Killing parameter, stamdfor the parameter corresponding to the point P with t = 0 being the parameter of Q. On the spacelike hypersurface E, choose arbitrary coordinates {xi} for Q. Assign the coordinates (t == xO, xl, x 2 , x 3 ) to P. Since F, does not chamge the metric, E, == F,E, the tramslation of E by F" is also orthogonal to the orbit of the isometry. Moreover, the components ofthe metric in this coordinate system cannot be dependent on the Killing parameter t. Thus, in this coordinate system, the spacetime metric takes the forru 3

9 = goodt ® dt -

L g'jdx

i

j ® dx .

(28.93)

i,j=l

spherically symmetric spacetimes

28.6.2. Definition. A spacetime is spherically symmetric if its isometry group contains a subgroup isomorphic to 80(3) and the orbits of this group are twodimensional spheres. In other words, if we think of isometries as the action of some abstract group, then this group must contain 80(3) as a subgroup. Since 80(3) is isomorphic to the group of rotations, we conclude that the metric should be rotationally invariamt. The time-tramslation Killing vector field must be orthogonal to the orbits of 80(3), because otherwise the generators of 80(3) camchamgethe projection of on the spheres amd destroy the rotational invariamce. Therefore, the 2-dimensional spheres must lie entirely in the hypersurfaces E,. Now, we cam write down a

e

e

924

28. DIFFERENTIAL GEOMETRY

s

u

Figure 28.2

t

Thecoordinates appropriate fora stationary spacetime.

static spherically symmetric metric in terms of appropriate coordinates as follows. Choose the sphetical coordinates (0, rp) for the 2-spheres, aod write the metric of this sphere as 92

= r 2dO I8i so + r 2 sin 2 Odrp I8i dtp,

where r is the ''radius'' of the 2-sphere. Choose the third spatial coordinate to be orthogonal to this sphere, i.e., r, Rotational symmetry now implies that the components of the metric must be independent of 0 aod rp. The final form of the metric based entirely on the assumed symmetries is 9 = f(r)dt I8i dt - h(r)dr I8i dr - r 2(dO I8i dO + sin 2 Odrp I8i dip).

(28.94) We have reduced the problem of finding ten unknown functions {g /LV }~, v=o offour

Schwarzschild solution ofEinstein's equation

variables {x/L }~~o to that of two functions f aod h of one variable. The remainiog task is to calculate the Ricci tensor corresponding to Equation (28.94), substitute it in Einstein's equation (with the RHS equal to zero), and solve the resulting differential equation for f aod h. We shall not pursue this further here, aod we refer the reader to textbooks on general relativity (see, for example, [Wald 84, pp. 121-124]). The final result is the so-called Schwarzschild solution, which is

9=

2M)dt I8i dt - (2M)-1 1 - --;:dr I8i dr (1 - --;:- r 2(dO I8i dO + sin 2 Odrp I8i drp),

(28.95)

where M is the total mass of the gravitating body, aod the natural units of GTR, in which G = 1 = c, have been used.

28.6 GENERAL THEORY OF RELATIVITY

925

A remarkable feature of the Schwarzschild solution is that the metric components have singularities at r = 2M and at r = O. It turns out that the first singularity is due to the choice of coordinates (analogous to the singularity at r = 0, = 0, n , and rp = 0 in spherical coordinates of JR3), while the second is a true singularity of spacetime. The first singularity occurs at the so-called Schwarzschild radius whose numerical value is given by

e

Schwarzschild radius

rs =

2GM --2- ""

c

M

3 - km, M0

where M0 = 2 x 1030 kg is the mass of the Sun. Therefore, for an ordinary body such as the Earth, planets, and a typical star, the Schwarzschild radius is well inside the body where the Schwarzschild solution is not applicable. If we relax the assumption of staticity, we get the following theorem (for a proof, see [Misn 73, p. 843]): 28.6.3. Theorem. (Birkhoff's theorem) The Schwarzschild solution is the only

spherically symmetricsolutionof Einstein'sequation in vacuum. A corollary to this theorem is that

28,6.4.Box. All spherically symmetric solutions of Einstein's equation in vacuum are static. This is analogous to the fact that the Coulomb solution is the ouly spherically syurmetric solution to Maxwell's equations in vacuum. It can be interpreted as the statement that in gravity, as in electromagnetism, there is no monopole "radiation."

28.6.3 Schwarzschild Geodesics With the metric available to us, we can, in principle, solve the geodesic equations [Equation (28.65)] and obtain the trajectories of freely falling particles. However, a more elegant way is to make further use ofthe syurmetries to eliminate variables. In particular, Proposition 28.4.4 is extremely useful in this endeavor. Consider first 9(80, u) where u is the 4-velocity (tangent to the geodesic). In the coordinates we are using, this yields 9(80, u)

= 9(80, x1L 81L ) = xlL 9(80, 81L), = r 20. r280IJ-

This quantity (because of Proposition 28.4.4 and the fact that 80 is a Killing vector field) is a constant of the motion, and its initial value will be an attribute of the

particle during itsentire motion. Weassignzerotothisconstant, i.e.,we assume that initially 0 = O. This is possible, because by rotating our spacetime-an allowed

926

28. DIFFERENTIAL GEOMETRY

operation due to rotational symmetry-we take the equatorial plane e = IT /2 to be the initial plane of the motion. Then the motion will be confined to this plane, because (j = 0 for all time. For the parameter of the geodesic equation, choose proper time r if the geodesic is timelike (massive particles), and any (affine) parameter if the geodesic is null (massless particles such as photons). Then gl'vil'iV = K, where K

= {I

o

for timelike geodesics, for null geodesics.

(28.96)

In terms of our chosen coordinates (with

e=

IT /2),

we have (28.97)

Next, we apply Proposition 28.4.4 to the time translation Killing vector and write (28.98)

e

where E is a constant of the motion and = at. In the case of massive particles, as r -+ 00, i.e., as we approach special relativity, E becomes i, which is the rest energy of a particle of unit mass.l'' Therefore, it is natural to interpret E for finite r as the total energy (including gravitational potential energy) per unit mass of a particle on a geodesic. Finally, the other rotational Killing vector field a~ gives another constant of motion, (28.99) which can be interpreted as the angnlar momentnm of the particle. This reduces to Kepler's second law: Equal areas are swept out in equal times, in the limit of Newtonian (or weak) gravity. However, in strong gravitational fields, spacetime is not Euclidean, and Equation (28.99) cannot be interpreted as "areas swept out." Nevertheless, it is interesting that the "form" of Kepler's second law does not change even in strong fields. Solving for i and


2

~t2 + ~ (1- 2M) (L2 +K) = ~E2 2

2

r

r

2

(28.100)

It follows from this equation that the radial motion of a particle on a geodesic is the same as that of a unit mass particle of energy E 2/2 in ordinary one-dimensional 13Recall thatthe -l-momentum of specialrelativity is p/k = mx/.L.

28.6 GENERAL THEORY OF RELATIVITY

927

t p

Figure 28.3

The effective potential VCr) for a massiveparticlewith L 2 = 5M 2.

nonrelativistic mechanics moving in aneffectivepotential 2

2

2

1 2M (L 1 M L ML ) V(r)=-(l--) -+1< =-1<-1<-+----. 2 r r2 2 r 2r2 r3

(28.101)

Once we solve Equation (28.100) for the radial motion in this effective potential, we can find the angular motion and the time coordinate change from (28.98) and (28.99). The new feature provided by GTR is that in the radial eqnation of motion, in addition to the "Newtonian term" -I< M/ r and the "centrifugal barrier" L 2 /2r 2 , we have the new term - M L 2 / r 3, which, a small correction for large r, will dominate over the centrifugal barrier term for small r. Let us consider first the massive particle case, I< = 1. The extrema of the effective potential are given by

(28.102) Thus, if L 2 < 12M 2 , no extrema exist (see Fignre 28.3), and a particle heading toward the center of attraction, will fall directly to the Schwarzschild radius r = 2M, the zero of the effective potential, and finally into the spacetime singularity r = a.For L 2 > 12M 2 , thereadermaycheckthatR+ isantinimumofV(r), while R_ is a maximum (Figure 28.4). It follows that stable (unstable) circular orbits exist at the radins r = R+ (r = R_). In the Newtonian limit of M « L, we get R+ '" L 2 / M, which agrees with the calculation in Newtonian gravity (Problem 28.30). Fnrthermore, Equation (28.102) puts a restriction of R+ > 6M on R+ and 3M < R_ < 6M on R_. This places the planets of the Sun safely in the region of stable circular orbits.

928

28. DIFFERENTIAL GEOMETRY

Ifa massive particle is displaced slightly from its stable equilibrium radius R+, it will oscillate radially with a frequency w, given by 14

W;=d2~1 dr

,~R+

=M;R+-6~). R+ - 3M R+

On the other hand, the orbital frequency is given by Equation (28.99), w2

~

2

M = _L = -..,,---~

Rt

R~-3MRr

where L 2 has been calculated from (28.102) and inserted in this equation. In the Newtonian limit of M « R+, we have w~ '" co, '" M/ R~. If w~ = w" the particle will return to a given value of r in exactly one orbital period and the orbit will close. The difference between w~ and w, in GTR means that the orbit will precess at a rate of

wp = w~ - w, = (1 - w,/w~)w~ = -[(1 - 6M/R+)1/2 - l]w~, which in the lintit of M

wp '"

« R+ reduces to

3M 3/2 3(GM)3/2 R 5/ 2 = 2R 5/ 2'

+

C

+

If we include the eccentricity e and denote the semimajor axis ofthe elliptical orbit by a, then the formula above becomes (see [Misn 73, p. 1110])

3(GM)3/2 wp '"

c2 (1 _ e 2)a5/ 2 '

(28.103)

Due to its proxintity to the Sun, Mercury shows the largest precession frequency, which, after taking into account all other effects such as pertorbations due to other planets, is 43 seconds of arc per century. This residual precession rate had been observed prior to the formulation of GTR and had been an unexplained mystery. Its explanation was one of the most dramatic early successes of GTR. We now consider the null geodesics. With K = 0 in Equation (28.101), the effective potential becomes L2

V(r) = 2r 3 (r - 2M),

(28.104)

which has a maximum at r = Rmax es 3M, as shown in Figure 28.5. It follows that in GTR, unstable circnlar orbits of photons exist at r = 3M, and that strong gravity has significant effect on the propagation of light. 14In theTaylor expansion of anypotential VCr) about the equilibrium position'u of a particle (of writ mass), it is the second derivative term that resembles Hooke'spotential, !kx 2 with k = (d 2 V /dr 2)ro.

28.6 GENERAL THEORY OF RELATIVITY

929

distance from origin, r

] --._-_ __ _+.---

~

,, ,

'_._._--:---".--

! --------·1-.-.-.-.-.-.--·---

...

-

'" _

Fi9ure 28.4

"0"

j.---

------.-.+. --"---

!

.o.

, ,

····T··········· .,i -

----_...._-_._---....1---

.

.

i

I

2M

4M

6M

i 8M

10M

The effective potential V(r) for a massiveparticle with L 2 = 20M2.

The minimum energy required to overcome the potential barrier (and avoid falling into the infinitely deep potential well) is given by

1E2

2

L2 = V(Rmax) = 54M2

=>

L2 2. E2 = 27M

In flat spacetime, L/ E is precisely the impact parameter b of a photon, i.e., the distance of closest approach to the origin. Thus the Schwarzschild geometry will capture any photon seuttoward it if the impact parameteris less than be es .fYiM.

Hence,the cross sectionfor photon capture is a sa :n:b~ = 27:n:M2.

To analyze the bending of light, we write Equation (28.100) as

~ (dr)2
2

Substituting for rp from Equation (28.99) and writing the resulting DE in terms of a new variable u = M/ r, we obtain d U) 2 b2 2(1-2u) ( drp +u = M2'

E

"
where we used Equation (28.104) for the effective potential. Differentiating this with respect to rp, we finally get the second-order DE,

d2 u -2 drp

+u =

3u 2

(28.105)

930

28. DIFFERENTIAL GEOMETRY

distance from origin, r f-

..

1\

\

I-

1-.... ..

'\i\

i\

1-...

f-.... ····-i--··········

i

i R Figure 28.5

,

r-;

10M

i

i

i

,

30M

,

i

,

,,

i

40M

The effective potential V (r) for a massless particle. Note that the shape of

the potential is independent of L 2 •

In the large-impact-parameter or small-u approximation, we can ignore the secondorder term on the RHS and solve for u. This will give the equation of a line in polar coordinates. Substituting this solution on the RHS of Equation (28.105) and solving the resulting equation yields the deviation from a straight line with a deflection angle of

4M

8


gravitational redshilt discussed

b

4GM =

(28.106)

bc2 '

where we have restored the G's and the c's in the last step. For a light ray grazing the Sun, b = R@ = 7 x lOSmandM = M@,= 2x 1030 kg, so that Equation (28.106) predicts a deflection of 1.747 seconds of arc. This bending of starlight passing near the Sun has been observed many times, beginning with the 1919 expedition led by Eddington. Because of the intrinsic difficulty of such measurements, these observations confirm GTR ouly to within 10% accuracy. However, the bending of radio waves emitted by quasars has been measured to an accuracy of I %, and the result has been shown to be in agreement with Equation (28.106) to within this accuracy. The last topic we want to discuss, a beautiful illustration of Proposition 28.4.4 is the gravitational redshift.Let 01 and 02 be two static observers (by which we mean that they each move on an integral curve of the Killing vector field e). It follows that the 4-velocities U1 and U2 of the observers are proportional to Since U1 and U2 have unit lengths, we have

e.

i = 1,2.

28.6 GENERAL THEORY OF RELATIVITY

931

distance from origin, r

1-...0.""."\-+ 1-.. .

F--. ...

;............

···············r

E--. .. .

Figure 28.6

-

, ..

.j

-"

+.................. ...,................. c

! ,

•• 000"._ ••• _------, •• ---

••••

.0.

j

_

i

i

i

5M

10M

15M

.

c

20M

A spacetime diagram depicting the emissionof lightby observer 01 andits

reception by Oz. Suppose Oi emits a light beam and Oz receives it (see Figure 28.6). Since light travels on a geodesic, Proposition 28.4.4 gives (28.107) where u is tangent to the light trajectory (or the light signal's 4-velocity). The frequency of light for any observer is the time component of its 4-velocity, and because the 4-velocity of an observer has the form (I, 0, 0, 0) in the frame of that

observer, we canwrite this invariantly as Wi = g(u, u.) =

g(n, ei), ,,!g(ei' ei)

i = 1,2.

In particular, using Equation (28.107), we obtain

1-2M/rz 1-2M/ri'

e)

where we used g(e, = gOO = (I - 2M/r) for the Schwarzschild spacetime, and ri and ri are the radial coordinates of the observers Oi and Oz, respectively. In terms of wavelengths, we have

1- 2M/,! 1- 2M/rz

(28.108)

It follows from Equation (28.108) that as light moves toward regions of weak gravity (rz > ri), the wavelength increases ('.Z > AO, i.e., it will be "red-shifted."

932

28. DIFFERENTIAL GEOMETRY

this makes sense, because an increase in distance from the center implies an increase in thegravitational potential energy. and, therefore, a decrease in a photon '8 energy luo. Pound and Rebka used the Mossbauer effect in 1960 to measure the change in the wavelength of a beam of light as it falls down a tower on the surface of the Earth. They found that, to within the 1% experimental accuracy, the GTR prediction of the gravitational redshift was in agreement with their measurement.

28.7 Problems 28.1. Show that de = 0 A (Ge) - e A (GO). 28.2. Let Aand 8 be matrices whose elements are one-forms. Show that (AA8)' = -8' AA'. 28.3. Write Equation (28.20) in component form and derive Equation (28.24). 28.4. Find dO if

0- ( -

0

cot OeqJ

-cote.~)

o

'

where (e, 'P) are coordinates on the unit sphere 8 2 . 28.5. Find the curvature of the two-dimensional space whose arc length is given by ds 2 = dx 2 + x 2dy 2. 28.6. Find the curvature of the three-dimensional space whose arc length is given by ds 2 = dx 2 + x 2dy 2 + dz 2 . 28.7. Find the curvature tensors of the Friedmann and Schwarzschild spaces given in Example 28.2.3. 28.8. (a) Show that in]R3 the composite operator do

* gives the curl of a vector

whenthe vectoris written as components of a two-form.

*

(b) Similarly, show that 0 d is the divergence operator for one-forms. (c) Use these results and the procedure of Example 28.2.4 to find expressions for the curl and the divergence of a vector in curvilinear coordinates. 28.9. Let 0 = er, v = eko and w =

.j

in (28.34) to arrive at (28.35).

28.10. Using Equations (28.7) and (28.55) show that R is a tensor of type (1, 3). In particular, it does not differentiate functions that multiply vectors and the l-form in its argument. 28.11. Show that R(o, v)w + R(w, o)v + R(v, w)o = O. Hint: Use the first property of Theorem 28.3.2 to change the vector with respect to which covariant differentiation is being performed. Then use the Jacobi identity for Lie brackets of vectors.

28.7 PROBLEMS

933

28.12. Use (28.55) and (28.11) to show that the components of R are precisely those introduced in Equation (28.22). 28.13. Prove the statement in Box 28.3.8. 28.14. Start with d 2x i /dt 2 = 0, the geodesic equations in Cartesian coordinates. Transform these equations to spherical coordinates (r, II, rp) using x = r sin II cos rp, y = r sin II sin rp, and z = r cos II, and the chain rule. From these equations read off the connection coefficients in spherical coordinates [refer to Equation (28.65)]. Now use Equation (28.43) and Definition 28.3.6 to evaluate the

divergence of a vector. 28.15. Find the geodesics of a manifold whose arc element is ds 2 = dx 2 + d y2 + dz 2 • 28.16. Find the geodesics of the metric ds 2 = dx 2 + x 2dy2. 28.17. Find the differential equation for the geodesics of the surface of a sphere of radius a having the line element ds 2 = a 2dll2 + a 2 sm2 IIdrp2. Verify that A cos rp + B sin rp + cot II = 0 is the intersection of a plane passing through the origin and the sphere. Also, show that it is a solution of the differential equation of the geodesics. Hence, the geodesics of a sphere are great circles. 28.18. The Riemann normal coordinates are given by xi = a i t. For each set of a i , one obtains a different set of geodesics. Thus, we can think of a i as the parameters that distinguish among the geodesics. (a) By keeping all a i (and t) fixed except the jth one and using the definition of tangent to a curve, show that nj = taj, (one of) the nCs) appearing in the equation of geodesic deviation. (b) Substitute (a) plus ui = xi = a i in show that R'ijk

+ R'jik =

3 (rki,,i

+ rkj,i)'

Substitute for one of the P's on the RHS using Equation \""O,.OUI). (c) Now use the cyclic property of the curvature that

r"!·I}, k =

tensor to

show

-~3 (R"!'k + Rn~'k)' JI I}

28.19. Let Vr : M --> N be isometric at injective. Hint: Look at the null space of Vr•. 28.20. (a) Show that the covariant and are related by (Luw, v) = (Vuw, v)

+ (w, Vvu) .

(b) Use this to derive the identity (LuW)i = (VuW)i

+ w,i (Ve i u),i .

Vr.

is necessarily

appuen to a l-form,

934

28. DIFFERENTIAL GEOMETRY

28.21. Show that Equation (28.69) leads to Equation (28.70). 28.22. Show that a vector field that generates a conformal transformation satisfies Xk8kgij

+ 8iXkgkj + 8jXkgki

=

-1{rgij.

28.23. (a) Show that R(u, v)(f) = 0 for all functions defined on M. (b) Using Equation (28.64) show that 0= R(u, v)(w· x) = w- [R(u, v)x] (c) Conclude from (b) that

Rijkl

=

+ [R(u, v)w]

. x.

-Rjikl.

28.24. Use the symmetries of Rijkl [Equations (28.23) and (28.24)] to show that Rijkl = Rklij and R[ijkll = O. 28.25.· Use the symmetry properties of Riemann curvature tensor to show that (a) R i i j k = R m . / = 0, and (b) Rij = Rji. (c) Show that Rijkl;i + Rjk;l - Rjl;k = 0, and conclude that V . G = 0, or, in component form,

or; =

O.

28.26. Show that in an n-dimensional manifold without metric the number of

independent components of the Riemann curvature tensor is

If the manifold has a metric, the number of components reduces to

28.27. (a) Take the trace of both sides ofR- !Rg =KTtoobtainR = -KT/: ea

-KT. (b) Use (a) to obtain Roo = !K(Too + Tj). (c) Now use the fact that in Newtonian limit Tij Too'" P to conclude that agreement of Einstein's and Newton's gravity demands that K = 8". in units in which the universal gravitational constant is unity.

«

28.28. (a) Show that the most general second-rank symmetric tensor Eij constructed from the metric and Riemann curvature tensors that is linear in the curva-

ture tensor is

where a, b, andA are constants.

-!a.

(b) Show that Eij has a vanishing divergence if and only if b = (c) Show that in addition, Eij vanishes in flat space-time if and only if A = O.

2B.7 PROBLEMS

935

28.29. Show that R+ and R_, as given by Equation (28.102) are, respectively, a minimum and a maximum of V (r). 28.30. Use F = ma to show that in a circular orbit of radius R, we have L 2 = GMR. 28.31. Show that R+ > 6M and 3M < R_ < 6M, where R± are giveu by Equation (28.102). 28.32. Calculate the energy of a circular orbit using Equation (28.100), and show that E(R) =

R-2M

../R2_3MR

,

whereR = R± 28.33. Show that the radial frequency of oscillatiou ofa massive particle in a stable orbit of radius R+ is given by

2 lV r

=

M(R+ -6M) R:j. _ 3MR+ .

28.34. Derive Equation (28.106) from Equation (28.105).

Additional Reading 1. Misner, C., Thome, K., and Wheeler, J. Gravitation, Freeman, 1973. A classic text on index-free differential geometry. This is the definitive text on Einstein's general theory of relativity. 2. Nakahara, M. Geometry,Topology, and Physics, Adam Hilger, 1990. A very readable book for graduate students of physics with a lot of emphasis on gauge theories and topological concepts. The book has a standard introduction to manifolds and a good discussion of differential geometry, especially Killing vector fields. 3. Wald, R. General Relativity, Uoiversity of Chicago Press, 1984. A more up-to-date book on differential geometry and general relativity.

29,

_

Lie Groups and Differential Equations

Lie groups and Lie algebras, because of their manifold-and therefore, differentiability-structure, find very natural applications in areas of physics and mathematics in which symmetry and differentiability play important roles. Lie himself started the subject by analyzing the symmetry of differential equations in the hope that a systematic method of solving them could be discovered. Later, Emmy Noether applied the same idea to variational problems involving symmetries and obtained one of the most beautiful pieces of mathematical physics: the relation between symmetries and conservation laws. More recently, generalizing the gauge invariance of electromagnetism, Yang and Mills have considered nonahelian gauge theories in which gauge invariance is governed by a nonahelian Lie group. Such theories have been successfully built for three of the four fundamental interactions: electromagnetism, weak nuclear, and strong nuclear. Furthermore, it has been possible to cast the fourth interaction, gravity-as described by Einstein's general theory of relativity-in a langnage very similar to the other three interactions with the promise ofnnifying all four interactions into a single force. This chapter is devoted to a treatment ofthe first topic, application ofLie groups to DEs. The second topic, the calculus ofvariations and conservation laws, will be discussed in the next chapter. The third topic, that of gauge theories, although of fundamental importance to the development of physics and our understanding of the universe, is, at this stage, too specialized to be covered in this book.

29.1 Symmetries of Algebraic Equations The symmetry group of a system ofDEs is a transformation group that acts on both the independent and dependent variables and transforms solutions of the system to

29.1 SYMMETRIES OF ALGEBRAIC EQUATIONS

937

other solutions. In order to understand this symmetry group, we shall first tackle the simpler question of the symmetriesof a system of algebraic equations. G-invariance and

symmetry group defined

29.1.1. Definition. Let G be a local Lie group of transformations acting on a manifold M. A subset ScM is called G-invariant and G is called a symmetry group of S if whenever g . P is defined for PES and g E G, then g . PES. 29.1.2. Example. Let M

= 1Il2 .

(a) Let G = lR+ be the abelian multiplicative group of real numbers. Let it act on M componentwise: r- (x, y) = (rx, ry). Then any line going throughthe origin is a G-invariant subset of]R2. (b) If G = 80(2) and it acts on M as usual, then any circle is a G-invariant subset of

.

~.

system of algebraic equations and their symmetry group

A system ofalgebraic equations is a system of equations Fv(x) = 0,

v=1,2, ... ,n,

io which F; : M --+ IR is smooth. A solution is a poiot x E M such that F v(x) = 0 for v = I, ... , n. The solution set of the system is the collection of all solutions. A Lie group G is called a symmetry group of the system if the solution set is G-iovariant. invariant map

29.1.3. Definition. Let G be a local Lie group of transformations acting on a manifold M. A map F : M --+ N, where N is another manifold, is called a G-invariant map iffor all P E M and all g E G such that g . P is defined, F (g. P) = F (P). A real-valued G-invariantfunction is called simply an invariant.

The crucial property of Lie group theory is that locally the group and its algebra "look alike." This allows the complicated nonlinear conditions of invariance

of subsets and functions to be replaced by the simpler lioear conditions of iovariance under infinitesimalactions.From Definition 27.1.25, we obtaio the following proposition. 29.1.4. Proposition. Let G be a local group oftransformations acting on a manifold M. A smooth real-valued function f : M --+ IR is G-invariant if and only if forall

P

E

M

andfor every infinitesimal generator

(29.1)

eE g.

29.1.5. Example. Theiofinitesimal generator for 80(2) is eM = x8 y - y8 x . Any functionoftheform f (x 2 + y2) is an 80 (2)-iovatiant. To seethis, weapply Proposition 29.1.4: (x8 y - y8x)f(x 2 + y2)

= x(2y)f' -

where /' is the derivative of f.

y(2x)!,

= 0,



938

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

The criterion for the invariance of the solution set of a system of equations is a little more complicated, because now we are not dealing with functions themselves, but with their solutions. The following theorem gives such a criterion (for a proof, see [Olve 86, pp. 82-83]):

29.1.6.Theorem. Let G be a local Lie group oftransformations acting on an m-

dimensional manifold M. Let F : M -> En, n :s m, define a system of algebraic equations {Fv(x) = O}~~l' and assume that the Jacobian matrix (8FvI8x k ) is of rank n at every solution x ofthe system. Then G is a symmetry group ofthe system ifand only if (29.2)

for every infinitesimal generator

e

E

D.

Note that Equation (29.2) is required to hold only for solutions x ofthe system.

29.1.7.Example. Let M

= 1R2 and G =

SO(2). Consider F : M -+ IR defined by

F(x, y) = x 2 + y2 - I. The Jacnbian matrix is simply the gradient, (8Flax, 8Flay) = (2x, 2y),

andis of rank 1 for all points of the solutionset, because it nevervanishes at the points where F(x, y) = O,i.e.,the nnit circle. It follows from Theorem 29.1.6that G is a symmetry group of the equation F(x, y) = 0 if and only if f;Mlr(F) = 0 whenever rESt. But f;Mlr(F)

= (xa y -

y8x)Flr

= 2xy -

2yx

= O.

Tbis is a proof of the obvious fact that SO (2) takes points of St to other points of S 1. As a less trivial example, consider thefunction F : ]R2 ---+ .IR given by

i

F(x, y) =x 2

+ y4+ 2x2 + y2 -2.

The infinitesimal actionof the group yields

The reader may check that f;M(F) = 0 whenever F(x, y) = O. The Jacobian matrix of the "system" of equations is thegradient

which vanishes only whenx = 0 = y. which doesnotbelongtothesolution set.Therefore, therank of the Jacobian matrix is 1. Weconclude that the solution set of F(x, y) = 0 is a rotationally invariant subset of]R2.Indeed, we have

andthesolution set is justthe unit circle.Notethatalthough thesolutionset of F (x, y) is G-invariant, the function itself is not.

=0 II1II

939

29.1 SYMMETRIES OF ALGEBRAIC EQUATIONS

We now discuss how to find invariants of a given group action. Start with a one-parameter group and write

.a ax'

v=eM=X'-. for the infirtitesimal generator of the group in some local coordinates. A local invariant F(x) of the group is a solution of the linear, homogeneous POE

aF

aF

ax

ax

= Xl (x)-l + ... + Xn(x)-n = o.

v(F)

(29.3)

It follows that the gradient of F is perpendicular to the vector v. Since the gradient of F is the normal to the hypersurface of constant F, we may consider the solution ofEquation (29.3) as a surface F(x) = c whose normal is perpendicular to v. Each normal determines one hypersurface, and since there are n -I linearly independent vectors perpendicular to v, there must he n - I different hypersurfaces that solve (29.3). Let us write these hypersurfaces as j = 1,2, ... , n - 1,

(29.4)

and note that .

6.FJ ""

aFi . Ln --. 6.x' = 0, i=l

j = 1,2, ... , n - 1.

ax'

A solution to this equation is suggested by (29.3): !::::. i _ x i

x

characteristic system

-CI!

i

;;;;}

!::::.x _ Xi -C\!.

For 6.x i ---> dx', we obtain the following set of ODEs, called the characteristic system of the original PDE,

ofa POE

dx l

dx 2

dx"

Xlix) = X2(x) = ... = xn(x)'

(29.5)

whose solutions detenmine {Fi (x)}j:J To find these solutions,

29.1.8. Box. Take the equalities of(29.5) one at a time, solve the first order DE, write the solution in the form of (29.4), and read off the functions.

The reader may check that any function of the F i's is also a solution of the PDE. 10 fact, it can he showo that any solution of the POE is a function of these F i 's (see [Olve 86, pp. 86-90]).

940

29" LIE GROUPS AND DIFFERENTIAL EQUATIONS

29.1.9. Example. Once again, let us consider SO(2), whose infinitesimalgenerator is -yax + xayo The characteristic "system" of equations is dx

-y Thus, F (x, y) = x 2 + y2, or any function thereof. is an invariant of the rotation group in two dimensions. As a less trivial example, consider the vector field

v=

-y~ +x~ +Ja 2 -z2~ ax

ay

az

where a is a constant.The characteristicsystem of ODEs is dx -y

dy

dz

= 7 = Ja2-z2"

Jx

2 + y2 = The first equation was solved above, giving the invariant Fl (x, y, z) = find the other invariant, solve for x and substitute in the second equation to obtain

r, To

dy dz Jr 2_ y2 = Ja 2-z2"

The solution to this DE is arcsin~ = arcsin~ +C r a '-..-' a p

=>

arcsin ~ - arcsin':' = C.

'-..-'

r

a

Hence, F2(X, y, z) = arcsin(yjr) -arcsin(zja) is a second invariant. By takingthe sine of Fl. we can come up with an invariant that is algebraic (rather than trigonometric) in form:

s = sin Fl = sin(a -

fJ)

= sin a cos fJ -

cos a sin fJ

=~.Jl_z22 _jl_y2~=y~-xz" 2 r

a

r a

ra

Any function of r and s is also an invariant.

II

When the dimension of the Lie group is larger than one, the computation of the invariants can be very complicated. If {vdk=1 form a basis for the infinitesimal generators, then the invariants are the joint solutions of the system of first order PDEs n . aF vk(F) = LX£(x)j' j=! ax

k = 1, .. .,».

To find such a solution, one solves the first equation and finds all its invariants. Since any function of these invariants is also an invariant, it is natural to express F as a function of the invariants of VI. One then writes the remaining equations in terms of these new variables and proceeds inductively.

29.2 SYMMETRY GROUPS OF OIFFERENTIAL EQUATIONS

941

29.1.10. Example. Considerthe vector fields u = -y8x + x8 y, y2z+a 3x) _ (X 3z +x 8 v~ x+ yX-+y-

(X 2yZ ~ +y3 Z+a3y) 8y _ (~+ 2 2 yX--j-y-z + b 3) 8e yX-+y-

where a and b are constants. The invariants ofu are functions of r = .,jxl + y2 and z. If we are to have a nontrivial solution, the invariant of v as well as its PDE should be expressible in tenus of r = xl + y2 and z. The reader may verify that

.J

with the characteristic equation

dr

dz

This is an exact first-order DE whose solutions are givenby

!r 2z2 + a3z + b 3r = c with c an arbitrary constant. Therefore, F

= !r2z2 + a 3z + b3r, or

F(x,y,z) = !(x2+y2)z2+a3z+b3Jx2+y2, III

or a function thereof, is the single invariant of this group.

29.2 Symmetry Groups of Differential Equations Let S be a system ofpartial differential equations iuvolving p independent variables x = (xl, ... , x P), and q dependent variables u = (u l, ... , u q ) . The solutions of the system are of the form u f(x), or, in component form, u" f(x l , ••• , x P ) , a = I, ... , q. Let X = ffi,P and U = ffi,q be the spaces of independent and dependent variables with coordinates {xi} and {u·}, respectively. Rougbly speaking, a symmetry group of the system S will be a local group of transformations that map solutions of S into solutions of S.

=

=

Marius Sephus Lie (1842-1899) was the youngest son of a Lutheran pastor in Norway. He studied mathematics and science at Christiania (which became Kristiania, then Oslo in 1925) University where he attended Sylow's lectures on group theory. There followed a few years when he could not decide what career to follow. He tutored privately after his graduation and even dabbled a bit in astronomy and mechanics. A turningpoint camein 1868 when he read paperson geometryby Poncelet and Plucker from which originated the inspiration in the topic of creating geometries by using elements other than points in space, and provided the seed for the rest of Lie's career, prompting him to call himself a student of Plucker, even though the two had never met. Lie's first

942

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

publicationwonhim a scholarshipto workin Berlin, where he met Klein, whohad also been influenced by Plucker's papers. The two had quite different styles-Lie always pursuing the broadest generalization, while Klein could become absorbed in a charming special case-but collaborated effectively for many years. However, in 1892 the lifelong friendship between Lie and Klein broke down, and the following year Lie publicly attacked Klein, saying, "1 am no pupil of Klein, nor is the opposite the case, although this might be closer to the truth." Lie and Kleinspenta summerin Paris, then parted for sometime beforeresuming their collaboration in Gennany. While in Paris, Lie discovered the contact transformation, which, for instance, maps lines into spheres. During the Franco-Prussian War, Lie decided to hike to Italy. On the way, however, he was arrested as a German spy and his mathematics notes were assumed to be coded messages. Only after the intervention of Darboux was Lie released, and he decided to return to Christiania. In 1871 Lie became an assistant at Christiania and obtained his doctorate. After a short stay in Germany, he againretumed to Christiania University, where a chair of mathematics was created for him. Several years later Lie succeeded Klein at Leipzig, where he was stricken with a condition, then called neurasthenia, resulting in fatigue and memory loss and once thought to result from exhaustion of the nervous system. Although treatment in a mental hospital nominally restored his health, the once robust and happy Lie became ill-tempered and suspicious, despite the recognition he received for his work. To lure him back to Norway, his friends at Christiania created another special chair for him, and Lie returned in the fall of 1898. He died of anemia a few months later. Lie had started examining partial differential equations, hoping that he could find a theory that was analogous to Galois's theory of equations. He examined his contact transformations considering how they affected a process due to Jacobi of generating further solutions from a given one. This led to combining the transformations in a way that Lie called a group, but is today called a Lie algebra. At this point he left his original intention of examining partial differential equations and examined Lie algebras. Killing was to examine Lie algebras quite independently of Lie, and Cartan was to publish the classification of semisimple Lie algebras in 1900. Much of the work on transformation groups for which Lie is best known was collected with the aid of a postdoctoral student sent to Christiania by Klein in 1884. The student, F. Engel, remained nine months with Lie and was instrumental in the production of the three volume work Theorie der Transformationsgruppen; which appeared between 1888 and 1893. A similar effort to collect Lie's work in contact transformations and partial differential equations was sidetracked as Lie's coworker, F. Hausdorff, pursued other topics. The transformation groups now known as Lie groups provided a very fertile area for research for decades to come, although perhaps not at first. When Killing tried to classify the simple Lie groups, Lie considered his efforts so poor that he admonished one of his departing students with these words: "Farewell, and if ever you meet that s.o.b., kill him." Lie's work was continued (somewhat in isolation) by Cartan, but it was the papers of Way! in the early 1920s that sparked the renewal of strong interest in Lie groups. Much of the foundationof the quantumtheory of fundamentalprocesses is built on Lie groups. In 1939,

29.2 SYMMETRY GROUPS OF DIFFERENTIAL EQUATIONS

943

Wigner showed that application of Lie algebras to theLorentz transformation required that all particles have theintrinsic properties of mass andspin.

transform the graph ofa functlon to find thefunction's transform!

To make precise the above statement, we have to clarify the meaning nf the action of G on a function u = f (x). We start with identifying the function f (i.e., a map) with its graph (see Chapter 0),

rl ea {(x, f(x»[x

E

Q} eX

x U,

where Q e X is the domain of definition of f. If the action of g E G on I'I is defined, then the transform of r I by g is

g. r 1= {(i, it) = g. (x, u) [(x, u) transform of a function by a group element

E

r I}'

In general, g . r I may not represent the graph of a function-in fact, it may not be even a function at all. However, by choosing g close to the identity of G and shrinking the size of n, we can ensure that g . r I = r J' i.e., that g . r I is indeed the graph of a function it = I (i). We write I = g . f and calli the transform of fby g. 29.2.1. Example. Let X = IR = U, so that we are dealing with an ODE. Let G = SO(2) be therotation group actingon X x U = JR2. The actionis given by

+ ucosO). f(x) is a function, the gronp SO(2) acts on its graph r I

(i, il) = O· (x, u) = (x coss - u sinO,x sinf

(29.6)

Ifu = by rotating it. This process canleadto a rotated graph () . I' i» whichmaynotbe thegraph of a single-valued function. However, if we restrict theinterval of definition of f. andmake (} smallenough, then () . I'f will be the graph of a well-defined function il = J(x) with r J = 0 . r I' If we substitnte f(x) for u, we obtain

(x, il) = 0 . (x, f(x)) = (x coes - f(x) sm s, x sinO + f(x) ccs s), or

x =xcosO -

f(x) sinO,

(29.7)

il=xsinO+f(x)cosO.

Eliminating x from thesetwo equations yields il in terms of i, from whichthe function j can be deduced. As a specific example,consider f(x) = kx 2. Then, the firstequation of (29.7) gives k ' 0) 2 ( sm x -cosf}x+x=O

=>

x=

cosO-Jcos20-4kisinO 2ksin8 '

where we kept the root of the quadratic equation that gives a finite answer in the limit

o --> D. Insertingthis in the secondequation of (29.7) and simplifyingyields _ f-(-) u= x

=

coss - Jcos20 - 4kx sinO _ 2 -xcotO. 2ksin 0

944

29. LIE GROUPS AND DIFFERENTIAL EQUATlDNS

Wewrite thisas _

!(x) ==

(0 .

f)(x) =

cos s - Jcos 20 - 4kx sinO 2

2k sin 0

This eqoation defines the function

1=

-

x cor s.

0 . I:

III

The foregoing example illustrates the general procedure for finding the transformed function I = g . f: 29.2.2. Box. If the rule oftransformation of g EGis given by

(i, it) = g. (x, u) = (Wg(x, u), g(X, u)), then the graph r j = g . r f of g . f is given parametrically by i = Wg(x, f(x)),

it = g(x, f(x)).

(29.8)

In principle, we can solve the first equation for x in terms of i and substitute in the second equation to find it in terms i, and consequently For some special but important cases, the transformed functions can be obtained explicitly. If G is projectable, i.e., if the action of G on x does not depend on u, then Equation (29.8) takes the special form i = Wg(x) and it = g(x, f(x)) in which 'I'g is a diffeomorphism of X with inverse If r f is the graph of a function f, then its transform g . r is always the graph of some function. In fact,

1-

projectable group

'1'.-"

it

= I(i) = g(x, f(x)) = g(W.-, (i),

f(W.-,(i))).

(29.9)

In particular, if G transforms only the independent variables, then it

= I(i) = f(x) = f(W.-,(i))

=}

1= f '1'.-"

For example, if G is the group of translations x will be defined by I(x) = f(x - a). symmetry group ofa system of DEs

0

f->

x

(29.10)

+ a, then the transform of f

29.2.3. Definition. A symmetry group of a system of DEs S is a local group of transformations G acting on an open subset M of X x U with the property that whenever u = f(x) is a solution ofS and I'" g. f is definedfor g E G, then u = I(x) is also a solution ofS. The importance of knowing the symmetry group of a system of DEs lies in the property that from one solution we may be able to obtain a family ofother solutions by applying the group elements to the given solution. To find such symmetry groups, we have to be able to "prolong" the action of a group to derivatives of the dependent variables as well. This is obvious because to test a symmetry, we have to substitute not only the transformed function u = I (x), but also its derivatives in the DE to verify that it satisfies the DE.

29.2 SYMMETRY GROUPS OF OIFFERENTIAL EQUATIONS

945

29.2.1 Prolongation of Functions Given a function f : X --+ JEt, there are

different derivatives of order k of f. We use the multi-index notation aJi(x) "'"

k

. a f(x)

.

aXil8xJ2 ... OXik

for these derivatives, where J = (h, ... , jk) is an unordered k-tuple of nonnegativeintegers with I :s: A :s: p (see also Section 21.1). The order of the multi-index J, denoted by IJ I, is the sum of its components and indicates the order of differentiation. So, in the derivative above, IJI = h + ... + A = k. For a smooth map f: X --+ U,wehavef(x) = (fl(x), ... , fq(x)), so tharwe necd q-ji, numbers a] f"(x) to represent all k-th order derivatives of all components of f. To geometrize the treatment of DEs (and thus facilitate the study of their invariance), we need to constmct a space in which derivatives of all orders up to a certain number n participate. Since derivatives need functions to act on, we arrive at the space of functions whose derivatives share certain common properties. To be specific, me make the following defiultion. n-equlvalenoe of functions

29.2.4. Definition. Let f and h be functions defined on a neighborhood of a point a E X with values in U. We say that f and hare n-equivalent at a if a] f" (a) = a]h" (a) for all '" and J corresponding to all derivatives up to nth order. The collection of all U -valued functions defined on a neighborhood of a will be denoted by r .(X x U), and all functions n-equivalent to f at a by j~ f. A convenient representative of such equivalent functions is the Taylor polynomial of order n (the terms in the Taylor series up to nth order) about a. Now collect all j~ f for all a and f, and denote the result by In (X x U), so that r(XxU)"",[j~fIVaEX

nth jet space of U

and VfEr.(XXU)).

(29.11)

reX x U) is called the nth prolongation of U, or the nth jet space of U. It tums out that r(X x U) is a manifold (see [Saun 89, pp. 98 and 199]). 29.2.5. Theorem. In(X x U) is a manifold with natural coordinate functions

(xi, u" , uj) defined by Xi(j~f)

Note that the "points" of J"(X x U) are U-valued funotions!

= ai,

u"(j~f)

= rea),

uj(j~f) = aJi"(a).

The natural coordinatefunctionsallowus to identifythe spaceof the derivatives with various powersof R, Let Uk "'" JEtqpk denote the set of coordinates uj, and let

946

29. LIE GROUPS ANO DIFFERENTIAL EQUATIONS

U(,,) '" U X UI X ... x U" be the Cartesian product space! whose coordinates represent all the derivatives u J of all orders from 0 to n, The dimension of U(,,) is

A typical point in U(,,) is denoted by u(,,), which has qp(") different components {uJ}~=I' where J runs over all unordered multi-indices J = (h,.··, ik) with 1 :0: ik :0: p and 0 :0: k :0: n. By convention, k = 0 refers to no derivatives at all, in which case we set J = 0 and sa u", The nth jet space J"(X x U) can now be identified with X x U(,,). From now on, we shall use X x U(,,) in place of J"(X xU).

ug

29.2.6. Example. Let p = 3 and u = 1, i.e., X = JR.3 and U = R. The coordinatesof X are (x, y, z) andthat of U is u. Thecoordinates of Ul are (UXI "y. uz), where thesubscript denotes thevariable of differentiation. Similarly, the coordinates of U2 are (uxx, Uxy. "xz. Uyy• Uyz. u zz)

and those of U(2)

== U

X UI

x

U2

are

(u; "x. "y. U z ; Uxx. Uxy. "xz. Uyy. uyz. u zz),

which shows that U(2) is lO-dimensional. prolongation of a function

III

29.2.7. Definition. Given a smooth map f : X :::J g -+ U, we define a map pr(") f : g -+ urn) whose components (pr(") f)~ are given by

This map is called the nth prolongation of f. Thus, for each x E X, pr(n) f(x) is a vector in jRqp(" whose components are the values of f and all its derivatives up to order n at the point x. For example, in the case of p = 3, q = 1 discussed above, pr(2) f(x, y, z) has components

When the underlying space is an open subset M of X xU, its corresponding jet space is M(n) sa M x UI

X •.• X

Un,

which is a subspace of X x urn) ~ J" (X x U). If the graph of f in M, the graph of pr(") f lies in M("). INote thatU, identified with the spaceof zeroth derivative, is a factor in

o».

:X

-+ U lies

29.2 SYMMETRY GROUPS OF OIFFERENTIAL EQUATIONS

947

Prolongation allows us to turn a system of DEs into a system of algebraic equations: Given a system of I DEs Ll. v ({xi}, {U U}, {aiUU}, {aiajUU}, ... , (ai"" ai,Uuj) = 0,

v = 1, ... , I,

one can define a map Ll. : M(n) -+ ~l and identify the system of DEs with

S" '" {(x,

u(n)

E M(n)

I Ll. (x, u(n)

= O}.

By identifying the system of DEs with the snbset S" of the jet space, we have translated the abstract relations among the derivatives of u into a geometrical object S", which is more amenable to symmetry operations. 29.2.8. Definition. Let Q be a subset of X and f : Q -+ U a smooth map. Then f is called a solution ofthe system ofDEs S" if Ll. (x, pln) f(x)) = 0

vx

E Q.

Just as we identified a function with its graph, we can identify the solution of a system of DEs with the graph of its prolongation pr(n)f. This graph, which is denoted by rr), will clearly be a subset of S,,:

rjn) '" {(x, prtn) f(x))} c s", 29.2.9. Box. An nth order system of differential equations is taken to be a subset S" of the jet space J"(X x U), and a solution to be a smooth map f : Q -+ J"(X x U) the graph of whose nth prolongation pr(n) f is contained in S",

29.2.10. Example. Consider Laplace'sequation

v 2u = u xx + Uyy + Uzz

= 0

= 3, q = I, and n = 2. The total jet space is the 13-dimensional Euclidean space X x U(2), whose coordinates aretaken to be

with p

(x, y, z; u; ux, u y. uz; u:cx, Uxy. uxz. Uyy. uYz. uzz)· In this 13-dimensional Euclidean space,Laplace's equation defines a 12-dimensional subspace g li. consisting of all points in the jet space whose eighth, eleventh, and thirteenth

coordinates add upto zero. A solution f : JR.3 :>

Q -+ U

V(X,y,Z)EQ.

c IR mustsatisfy

948

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

This is the same as requiring the graph I'

fer)

==

j) to lie in S 6.. For example, if

f(x, y, z) = x 3yZ - y3 x Z + y3 - 3YZ2,

then, collecting each section of pr(2) f with fixed a (see Definition 29.2.7) separately, we

have (pr(2) f) 1(r) = x 3yZ _ y3XZ + y3 _ 3YZ2, (pr(2) f)2(r) = (3x 2yz _ y3 Z,x 3Z - 3lxz + 3y2 _ 3z 2,x3y _ y3 x - 6yz) , (pr(2) f)\r) = (6xyz, 3x 2z - 3y2 Z, 3x 2y - y3, -6xyz

+ 6y, x 3 -

3xl- fiz, -6y)

whichlies in S.6. becausethe sumof the eighth,the eleventh, and the thirteenth coordinates of (x, y, z; pr(2) f(x, y, z)) is.tixyz - 6xyz + 6y - 6y = O. III

29.2.2 Prolongation of Groups

nth prolongation ofa group action

Suppose G is a group of transformations acting on M C X xU. It is possible to prolong this action to the n-jet space M(n). The resulting gronp that acts on M(n) is called the nth prolongation of G and denoted by pr(n)G with group elements pr(n)g, for g E G. This prolongation is defined naturally: The derivatives of a function f with respect to x are transformed into derivatives of j = g . f with respect to x. More precisely, pr(n) g .

U; /) '" (x, pr(n) j(x)) '" (x, pr(n) (g . /)(1')) .

(29.12)

For n = 0, Equation (29.12) reduces to the action of G on M as given by Equation (29.8). That the outcome of the action of the prolongation of G in the defining equation (29.12) is independent of the representative function f follows from the chain rule and the fact that only derivatives up to the nth order are involved. The following example illustrates this. 29.2.11. Example. Let x, U, and G be as in Example 29.2.1. In this case, we have pr(llO.

til f) es pr(l)o . (x, u, Ut) es (i, ii, iit).

We calculated x and

uin that example. They are

x=xcos6-usin8, u=xsin6+ucosB.

(29.13)

To find u1. we need to differentiate the second equation with respect to x and express the result in terms of the original variables. Thus

_ dii u1 ea di

=

dii dx dx di

( .

du

) dx. dx di = (smO + ul cosO) di:

= smO + dx cosO

dx / di is obtainedby differeotiating the firsteqoation of (29.13): ~

~

x

x

1= d- cosB - d- sinB =

~

~~

dx

dx dx

---=- cos s - ----;; sinO

~

= (cosO - Ul sinO)---;;.

dx

29.2 SYMMETRY GROUPS OF OIFFERENTIAL EQUATIONS

949

or

dx di = COS(}-Ulsin(}'

It therefore follows that _

u1

sin(}+UICOS(} = --;;--'---:----;: CDSe - Ul sine

and pr

(1) f),(x,u,ul)= ( xcos(}-usin01xsm(}+ucos(}, . OO8+U 1CO'8) 8 . 8 . cos -"1 sm

Wenote that theRHS involves derivatives up to order one. Therefore, the transformation is independent of the representative function. So, if we had chosen ilk where h is 1equivalent to I, we would have obtained the sameresult. Thisholds for derivatives of all orders. Therefore, theprolongation of theaction of thegroup G is well-defined. 11II

In many case" it is convenient to choose the nth-order Taylor polynomial as the representative of the class of n-equivalent functions, and, if possible, write the transformed function f explicitly in terms of i, and differentiate it to obtain the transformed derivatives (see Problem 29.3). Example 29.2.11 illustrates an important property of the prolongation of G. We note that the first prolongation pr(1)G acts on the original coordinates (x, u) in exactly the same way that G does. This holds in general:

29.2.12. Box. The effect of the nth prolongation prCnlG to derivatives up to order m :::: n is exactly the same as the effect ofpr(m)G. Ifwe already know the action ofthe mth-order prolonged group pr(mlG, then to compute prCnlG we need only find how the derivatives u~ oforder higher than m transform, because the lower-order action is already determined.

29.2.3

Prolongation of Vector Fields

The geometrization of a system of DEs makes it possible to use the machinery of differentiable manifolds, Lie groups, and Lie algebras to unravel the synunetries of the system. At the heart of this machinery are the infinitesimal transformations, which are directly connected to vector fields. Therefore, it is necessary to find out how a vector field defined in M C X x U is prolonged. The most natoral way to prolong a vector field is to prolong its integral curve-which is a one-parameter group of transformations of M-to a curve in MCnl and then calculate the tangent to the latter curve.

950

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

nth prolongation of a vector field

29.2.13. Definition. Let M be an open subset of X x U and X E X(M). The nth prolongation of X, denoted by pr(u)X, is a vector field on the nth jet space M(n) defined by

pr(n)XI(x u("» = ~pr(n)[exp(tX)l' (x, u(n»1 'dt ,~O for any (x, u(n»

E

M(n).

Since (x, u(n» E M(n) form a coordinate system on M(n), any vector field in M(n) can be written as a linear combination of a/ax i and a/au] with coefficients being, in general, functions of all coordinates xi and u]. For a prolonged vector, however, we have

a LX axa + LLX]-., au] P

pr(n)x=

i-

i=l

q

i

(29.14)

a=l J

o

wbere Xi and X are functions only of x k and u. This is due to the remark made in Box 29.2.12. For the same reason, the coefficients X] corresponding to derivatives of order m will be independent of coordinates u] that involve derivatives of order higher than m. Thus, it is possible to construct various prolongations of a given vector field recursively. 29.2.14. Example. Let us consider our recurrent example of X;: U ;: JR, G = SO(2). Given the infinitesimal generator v = curve to obtain2

-uax + xau , one can solve the DE of its integral

exp(tv)(x, u) = (x cost - u sint, x sint

+ u cost).

Example 29.2.11 calculated the first proloogation of SO (2). So

exp(tv),(x,U,Ut)= ( xcost-usint,xsint+ucost, Sint+UICOSl) .' cost -"1 smr Differentiating the components with respect to t at t = 0 gives pr

(1)

i(XCost-usint)1

at

1=0

=-u,

~(XSint+ucost)1 =x, at 1=0 -a (Sint+UICOSl)1 .

at

cost -"1 sm r

t=O

2 =I+u.

1

Therefore, pr(l)v =

a a a -u- +x- + (1 +uI)-. ax au aU!

Note that the first two terms in pr(l)v are the same as those in v itself, in agreement with ..

Box 29.2.12.

20ne can, of course, also write the finite group element directly.

29.3 THE CENTRAL THEOREMS

951

29.3 The Central Theorems We are now in a position to stale the first central theorem of the application of Lie gronps to the solution of DEs. This theorem is the exact replica of Theorem 29.1.6 in the language of prolongations:

29.3.1. Theorem. Let

v

=

1, ... ,1,

be a system ofDEs defined on M C X x U whose Jacobian matrix Ja(x,u

Cn) )= _

v 8A v ) (aA --. '-a-

ax' au]

has rank I for all (x, u(n») E SD.. JfG is a local Lie group oftransformations acting onM,and v for every infinitesimal generator system.

= I, ... , I,

whenever

A (x, u Cn»)

=0

eof G, then G is the symmetry group of the

29.3.2. Example. Consider the first order (so, n = I) ordinary DE A(x,u, Uj)

so that X

~

= (u _x 3 -

JR and U

~

u 2x)Uj +x +x 2u +u 3

= 0,

JR. We first note that

a A aA aA)

J.6,

= ( B;' a;;'

aUl

= «-3x 2 _ u 2)Uj + I

+ 2xu, (I -

2ux)Uj

+ x 2 + 3u 2, U _

x 3 _ u 2x),

which is of rank 1 everywhere. Now let us apply the first prolongation of thegenerator of SO (2)---<:a1culated in Example 29.2.14-to A. We have pr(l)v(A)

aA aA = - uaA - + x - + (l +UT)ax au aUj = -U[(-3x2 - u 2)Uj

+ I + 2xu] + x[(1 -

2ux)Uj

+ x 2 + 3u 2]

x 3 - u 2x) = Uj[(u - x 3 - u 2x)Uj +x +x 2u + u 3] = ujA.

+ (I + uT)(u -

It follows that pr(l)v(D.) = 0 whenever A = 0, and that SO(2) is a symmetrygroup of the DE.Thus, rotations will change solutions of theDE intoother solutions. In fact,thereader mayverifythat in polar coordinates, the DE canbe written in the incredibly simple form dr 3 dB =r .

andthe symmetry of the DE conveys the factthat adding a constant to (} does not change the polar form of the DE. II

952

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

Theorem 29.3.1 reduces the invariance of a system of DEs to a criterion involving the prolongation of the infinitesimal generators of the symmetry group. The urgent task in front of us is therefore to construct an explicit formula for the prolongation of a vector field. In order to gain insight into this construction, we first look at the simpler case of U ~ IR and a group G that transforms only the independent variables. Furthermore, we restrict ourselves to the first prolongation. An infinitesimal generator of such a group will be of the form p

a

.

v= LX'(x)-., l

ax

i=l

which is assumed to act on the space M C X xu. The integral curve of this vector field is exp(tv) which acts on points of M as follows: (E, it) = exp(tv)· (x, u) sa (W,(x), u) es (W;(x), u).

By the construction of the integral curves in general, we have dW;(x)

I

= Xi (x).

(29.15)

,~O

dt

Denote the coordinates of the first jet space M(ll by (xi, u, Uk), where aflax k . By the definition of the action of the prolonged group,

Uk(j; f) sa

pr(l) exp(tv) .

(j; f) = (W, (x), u, it [),

(29.16)

whereu = !(x)anditj = a!/ai j . Once we find itj, we can differentiate Equation (29.16) with respect to t at t = 0 to obtain the prolonged vector field. Using Equation (29.10) and commas to indicate differentiation,3 we obtain

Since v does not have any component along a/au, its prolongation will not have such components either. The components Uj along a/aUj are obtained by differ3Wehave found it exceedinglyconvenient to use commas to indicate differentiation with respectto the x's. The alternative,

i.e., the use of partials, makesit almostimpossible to findone's way in the maze of derivatives involvingxi • ii, and t, with:xi depending on t. The reader will recall that theindexafter the commais to be thought of as a "position holder," andtheargument as a SUbstitution. Thus,forexample, alesl,'''' sp) OSj

I S=X

~

alcol,,,., Op) I aOj

O='x

.

953

29.3 THE CENTRAL THEOREMS

entiating it j with respect to t:

.'!-\jI~'.j(X)1 dt ,~O

Uj(x, u, Uj) = tUi i=1

a~~',j (X)I

= tUi l=l

(29.17)

+ tUi 1=0

1=1

i: (\jI~'.jk(X)

a:tk)l_ .

k=l

/_0

The derivative of the first term in the sum can be evaluated as follows:

a\jl~"j _ -a-t-(x(t))

I

a\jl~,j

_

= - ---a:;-(x(-s))

I

(=0

a\jl:j sa - ---a:;-(s, x(-s))

I

8=0

=_

8=0

=_

a\jl~,j (X)I

(a\jl~ I )

a\jl:j as

8=0

= _

(S,x)1

a\jl:j (S,X(O))I as .

as

= _

as

8=0

,=0

8=0

.

J

(x) = _xi,(x), ,J

where we have emphasized the dependence of x on t (or s), treated s as the first independent variable, and in the second line substituted s = 0 in all x's before differentiation. This is possible because we are taking the partial derivative with respect to the first variable holding all others constant. The derivative of the second term in Equation (29.17) can be calculated similarly:

.

. \jib

a

2xi \jI'-, ,J'k(x(t))l ~, -0 = J'k (x (0)) = ' -k = 0 axJax

because

\jib = xi. We therefore have

E axi u; ax P

Uj(x, u, Uk) = -

(29.18)

-j

i=l

and k

a a ax E(x i -ax'a . +Ui--) =v- E -. Uk-· ax' aUi P

pr(l)v=

'-I

l_

P

(29.19)

. k-I t ,-

8Uj

It is also instructive to consider the case in which U is still JR, but G acts ouly on the dependent variable. Then v = U(x, u)a u , and (x, it) = (x, ,(x, u)),

.

WIth

U(x, u) =

d,(x, u) d t

I 1=0

.

The reader may check that in this case, the prolongation of v is given by 1

au

au

U·(x u ( » ) - - + u · J' - ax j J au .

(29.20)

954

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

The second equation in (29.20) can also be written as

Uii», ]

pr(l)

f(x»

=

au ax]

+ au

af au ax]

= ~[U(x, f(x»]. ax]

In other words, Ujt», u(l» is obtained from U(x, u) by differentiation with respect to xl , while treating u as a function of x. This leads us to the definition of the total derivative. total derivative

29.3.3. Definition. Let S : M(n) -> ffi. be a smooth function of x, u, and all derivatives of u up to nth order. The total derivative of S with respect to xi, denoted by DiS, is a smooth function DiS: M(n+l) -> ffi. defined by

DiS(j;+l f) =

a: i [S

(x, pr(n) f(x))];

i.e., DiS is obtainedfrom S by differentiating S with respect to xi, treating u and all the u'j's as functions ofx. The following proposition, whose proof is a straightforward application of the chaio rule, gives the explicit formula for calculating the total derivative: 29.3.4. Proposition. The ith total derivative of S : M(n) -> ffi. is ofthe form

as + ~"a as L-L-uJi-a

DiS = -,

ax'

'au]

a=l J

where,for J = (h,···, A),

and the sum over J includes derivatives ofall orders from 0 to n. An immediate consequence of this proposition is a

DiU]

au] = u]a . = -. ax ,I

'

Di(ST) = TDiS+SDiT.

V J, a, (29.21)

Higher-order total derivatives are defined in analogy with partial derivatives: If I is a multi-index of the form I = (ii, .. " ik), then the lth total derivative is (29.22) As in the case of partial derivatives, the order of differentiation is immaterial. We are now ready to state the second central theorem of the application of Lie groups to the solutiou of DEs (for a proof, see [Olve 86, pp. 113-115]).

955

29.3 THE CENTRAL THEOREMS

29.3.5. Theorem. Let

a

p.

v= L:x'(x,u)-.l i=1 CJx

q a + L:ua(x,u)-----;x a=l au

be a vector field on an open subset Me X x U. The nth prolongation of v, i.e., pr(n)v E X(M(n)), is

where for J = (h, ... , A), the inner sum extends over 1 :0: IJI :0: n and the coefficients U'j are given by

and the higher-order derivative D] is as given in Equation (29.22). 29.3.6. Example. Let p = 2, q = I, and consider the case in which G acts only on the independent variables (x, y). Thegeneral vector fieldforthis situation is

a

a

v=~(x'Y)ax +~(x'Y)ay'

Weare interested in the firstprolongation of this vector field. Thus,n = 1,X I = I;,X 2 = 11. and J has only one component, which we denote by i (also written as x or y). Theorem 29.3.5 gives 2

uJ sa Uj = -Dj EX i=l

andusingthenotation Ux pr(l)v

. au

l

-

ax j

2

i=l

2 ax au L j-j' i=1 ax ax j

. au,

+ L:xl----f =

ax

-

= au/ax andu y = au/ay, we obtain

a +U - a , = v + Ue-x--: y 8ux ClUy

(29.23)

where

U» =

a~

a~

ax

ax

-ux- -Uy-

In particular, if G

and

= 80(2), so that V =

-yax

+ xa y, then Ux =

-u y and Uy

=

ux. It

then follows that t

a +xa - U ya- +u a x-.

pre )v= - y -

ax

oy

aux

Buy

III

956

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

29.3.7. Example. Let p = I, q = I, and G = SO(2). The generat vector fieldfor this situation is

a

a

ax

au

V=-U-+X-.

For the first prolongation of this vector field n = 1, X l component, which we denote by x. Theorem 29.3.5 gives 1

a

I

UJ=Ux=Dx(x-X ux)+X uxx=l-

-u, and J has only one

I ax ax ux=l+u x·2

It follows that

a

I)

a

2a

pr( v=-u-+x-+(I+u ) - ,

ax

au

x

aux

whichis theresultobtained in Example 29.2.14. The secondprolongation can be obtained as well. Once againwe use Theorem 29.3.5 withobviouschangeof notation: Uxx

= DxDx(x - Xlux) + X1uxxx

=

Dx(l + u;

+ uuxx) -

UUxxx

= 3uxuxx.

Then

(2)

a

a

2a

a

pr v=-u-+x-+(I+ux)-+3uxuxx--. ax au aux auxx UsingTheorem 29.3.1, we note that the DE Ux x = 0 has SO(2) as a symmetry gronp, because with ..6.(x, U, ux. uxx) (2)

pr v(Ll.)

== "xx,

all. aA = - uaA- +x+ (I +u2x )aA- +3uxuxx-= 3uxuxx,

ax

au

aux

au xx

which vanishes whenever 6.(x, U, u x, u xx) vanishes. This is the statement thatrotations take straightlines to straighllines. III

29.4

defining equations forthe symmetry group ofa system of PDEs

Application to Some Known PDEs

We have all the tools at our disposal to compute (in principle) the most general symmetry group of almost any system of PDEs. The coefficients U'j of the prolonged vector field pr(n)v will be functions of the partial derivatives of the coefficients Xi and U" of v with respect to both x and u. The infinitesimal criterion of invariance as given in Theorem 29.3.1 will involve x, u, and the derivatives of u with respect to x, as well as Xi and U" and their partial derivatives with respect to x and u. Using the system of PDEs, we can obtain some of the derivatives of the u's in terms of the others. Substituting these relations in the equation of inlinitesimal criterion. we get an equation involving u's and powers of its derivatives that are to be treated as independent. We then equate the coefficients of these powers of partial derivatives of u to zero. This will result in a large numher of elementary

29.4 APPLICATION TO SOME KNOWN POES

957

PDEs for the coefficient functions Xi and U" of v, called the defining eqnations for the symmetry group of the given system of PDEs. In most applications, these defining equations can be solved, and the general solution will determine the most general infinitesimal symmetry of the system. The symmetry group itself can then be calculated by exponentiation of the vector fields, i.e., by finding their integral curves. In the remaining part of this section, we construct the symmetry groups of the heat and the wave equations.

29.4.1 The Heat Equation The one-dimensional heat eqnation u, = u xx corresponds to p = 2, q = I, and n = 2. So it is determined by the vanishing of .0. (x, t, u(2)) = u, - U xx . The most general infinitesimal generator of symmetry appropriate for this equation can be

written as

a

v = ;(x, t, u)ax

a

a

+ f(X, t, u)+ (x, t, u)-au ' at

(29.24)

which, as the reader may check (see Problem 29.11), has a second prolongation of the form

a

a

a

a

a

pr(2)v = v+x_ +'_ + xx_._ + xt __ + "-, 8ux aUt 8uxx 8uxt aUtt

where, for example, ' = , - ;,u x + (u - f,)U, - ;uuxu, - fuU~ xx = xx

+ (2xu -

;xx)u x - fxxU,

+ (uu -

2;xu)u~

- 2fxuUxU, - ;uuu~ - fuuU~U, (29.25) + (r./Ju - 2~x)uxx - 2-cx u x t - 3~uuxuxx - t'uUtUxx - 2'l'uuxuxt.

and subscripts indicate partial derivatives. Theorem 29.3.1 now gives

whenever

Ut

(29.26)

= U xx

as the infinitesimal criterion. Substituting (29.25) in (29.26), replacing u, with U xx in the resulting equation, and equating to zero the coefficients of the monomials in derivatives of u, we obtain a number of equations involving s, r, and . These equations as well as the monomials of which they are coefficients are given in Table 29.1. Complicated as the defining equations may look, they are fairly easy to solve. From (d) and (f) we conclude that t: is a function of t ouly. Then (c) shows that; is independent of u, and (e) gives 2;x f" or ;Cx, t) ~f,X + ~(t), for some arbitrary function ~. From (h) and the fact that; is independent of U we get uu = 0, or

=

(x, t, u) = £l(x, t)u

+ fJ(x, t)

=

958

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

Monomial

Coefficient Equation

2 u xx

0=0

2 uxu xx

7:uu

2~u +2,xu = 0 2,u = 0 2~x + 'xx - 't = 0 2,x = 0

UxU x x

UxUxt Uxx

u x' u x3 u x2 Ux I

Table 29.1 to them.

=0

~uu = 0 2~xu - tPuu = 0 ~xx -2tPxu -~t

tP,-tPxx=O

= 0

(a) (b) (c)

(d) (e)

(0 (g) (h) (i) (j)

Thedefining equations of theheatequation andthemonomials that give rise

for some as yet undetermined functions a and and (i) yields ~t = -2tPxu = -2a" or

fJ. Since ~ is linear in x,

~xx

= 0,

Finally, with tPt = a.u + fJt and tPxx = axxu + fJxx (recall that when taking partial derivatives, u is considered independent of x and t), the last defining equation (j) gives at = a xx and fJ, = fJxx, i.e., that a and fJ are to satisfy the heat equation. Substituting a in the heat equation, we obtain 1

-g'Lttt X

2

-

1

J.1'JttX

1 + Pt = -:;:(ttt.

which must hold for all x and t. Therefore, "Cttt

= 0,



= 0,

_ Pt -

1 -47:tt o

These equations have the solution ry = cst +C6.

It follows that a(x, t) = -~Clx2 - ~csx - ~CJt

+ C4 and

~(x, t) = ~(2CJt + C2)X + cst + C6, ,(t) = CJt2 + C2t + C3, tP(x, t, u) = (_~CJx2 - ~csx - ~CJt

+ C4)U + fJ(x, z),

959

29.4 APPLICATION TO SOME KNOWN POES

Inserting in Equation (29.24) yields I

v = [2 (2cII

+ cz)x + cst + C6] -axa + (cII z + czl + C3) -aaI

+ [( -!CIX Z -

!csx - !CJI

+ C4) u + fJ(x, I)] aau

= CJ [XI~

1 + IZ~ + Cz (!x~ + t~) ax al - -4(21 + x2)u~] au ax al a a +C3 +C4U~+CS(I~-!xuaa )+C6 +fJ(X,I)~. au ax u ax au at

Thus the Lie algebra of the infinitesimal symmetries of the heat equation is spanned by the six vector fields VI =

ax,

vs = 21ax

V2

-

=

at,

x8 x + 2t8t , 2 v6 = 41xax + 41 at - (x 2 + 21)uau

V3 =

xuau ,

u8 u,

(29.27)

V4 =

and the infinite-dimensional subalgebra vp = fJ(x, t)au , where fJ is an arbitrary solution of the heat equation. The one-parameter groups Gi generated by the Vi can be found by solving the appropriate DEs for the integral curves. We show a sample calculation and leave the rest of the computation to the reader. Consider Vs,whose integral curve is given by the set of DEs

dx -=2t, ds

dl = 0, ds

du -=-xu. ds

The second equation shows that I is not affected by the group. So, t = 10, where 10 is the initial value of I. The first equation now gives

dx

-ds = 210

=}

x

= 210s + xo,

and the last equation yields ~

-

ds

= -(210s

+ xo)u

=}

~ -

u

= -(210s

Changing the transformed coordinates to

+ xo)ds

2

=}

u = uoe-to' -Xo'

xi and removing the subscript from the

initial coordinates, we can write exp(vss) . (x, I, u)

= (x, i, ii) = (x + 21s, I, ue-.x-, 2').

Table 29.2 gives the result of the action of exp(vis) to (x, I, u).

960

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

Group Elemeut

Trausformed Coordinates (x, I, ii)

Gl = expfvj s) G2 = exp(v2S) G3 = exp(v3S) G4 = exptvas) G5 = exp(v5S)

(x+s,t,u) (x, t +s, u) (x, t, eSu) (eSx, e2St , u) 2 (x +2ts, I, ue-sx - s t)

2

G6 = exp(v6S) Gp = exp(vps)

Table 29.2

t - uJI-4stexp [ --sx X - --- -]) I - 4st' I - 4st ' I - 4st (x, t, u + sfJ(x, t)) (

Thetransformations caused by the symmetry group of theheatequation.

The symmetry groups Gl aud G2 reflect the invariauce of the heat equation under space aud time translations, G3 aud G p demonstrate the linearity of the heat equation: We cau multiply solutions by constauts aud add solutions to get new solutions. The scaling symmetry is contained in G 4, which shows that if you scale time by the square of the scaling of x, you obtain a new solution. G5 is a Galileau boost to a moving frame. Finally, G6 is a trausformation that cannot be obtained from auy physical principle. Since each group G; is a one-parameter group of symmetties, if f is a solution of the heat equation, so are the functions f; == G j • f for all i, These functions can be obtained from Equation (29.8). As au illustration, we find f6. First note that for u = f (x, r), we have

_

x

_

X=---,

1- 4st

t

t=---,

1- 4st

2

ii es l(x, t) = f(x, OJI - 4st exp [ -sx

1- 4st

] .

Next solve the first two equations above for x and t in terms of x and I:

x

x=---

1

1+ 4st'

t=---. I +4s1

Finally, substitute in the last equation to get

-

f(x, t) = f

(X I) rrr: [-sx 2 ] 1+ 4st' 1+ 4s1 VJ+"4sI exp 1+ 4s1

'

or, chauging x to x aud 1 to t,

2

I [ -sx ] f6(X, t) = JI +4st exp I +4st f

(X t) I +4st' I +4st .

29.4 APPLICATION TO SOME KNOWN POES

961

The other transformed functions can be obtained similarly. We simply list these functions:

= f(x - s, r), f3(x, r) = e' f(x, t),

b(x, t)

= f(x, t - s), f4(X, t) = f(e-'x, e-Z't), !2(x, r)

(29.28)

2st, t), fp(x, t) = f(x, t) + sf3(x, z), z f6(X,t) = 1 exp[ _sx ]f(_X_,_t_) . ../1 +4st 1 +4st 1 +4st 1 +4st

= e-,x+,2 t f(x -

fs(x, r)

We can find the fundamental solution to the heat equation very simply as follows. Let f (x, t) be the trivial constant solution c. Then

u = f6(X, t) =

c

../1 + 4st

e-,x'/(I+4st)

is also a solution. Now choose c = ../s/" and translate t to t - 1/4s (an allowed operation due to the invariance of the heat equation under time translation Gz). The result is

which is the fundamental solution of the heat equation [see (22.45)].

29.4.2

The Wave Equation

As the next example of the application of Lie groups to differential equations, we

consider the waveequation in two dimensions. Thisequation is written as Utt -

U xx

-

U yy

= 0, or

1Jij Uij

=0

and

.6. = rl i Uij,

(29.29)

where n = diag(l, -1, -1), and subscripts indicate derivatives with respect to coordinate functions xl = t, x Z = x, and x 3 = y. With p = 3 and q = 1, a typical generator of symmetry will be of the form

V=

3 . a a Lx(')-. +U-, i=l

ax'

au

(29.30)

where {X(i)l;=l and U are functions oft, x, y, and u to be determined. The second prolongation of such a vector field is

962

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

where by Theorem 29.3.5, U(i)

= o, (U -

tXCk)Uk)

+ tX(k)Uik = DiU -

k~l

UCijl = DiDj (U - tXCklUk) k=l

(DiX(k»Uko

+ tXCklUijk k~l

3

= DiDjU -

i:

k=1

k~l

L [(DiDjXCk»

Uk + Uik (DjXCk»

+ Ujk (Di X Ckl)],

k~l

and we have used Equation (29.21). Using (29.21) further, the reader may show that U(ijl = Uij

+ Uk (8jkUiU + 8ikUju

- xt»)

+ UlUk (8il8jkUuu - Xi:l8j, - Xj~8il) + Ukl (8il8jkUu - xik)8jl - xjk)8il) - UkUlm (X~kl8il8jm

(29.31)

+ X~ml8il8 jk + X~ml8ik8jl) -

UiU jUkX~~,

where a sum over repeated indices is understood. Applying pr(2)vto /1, we obtain the infinitesimal criterion U(tt) = UCxx)

+ U(yyl

or

ryi j U(ij) = O.

Multiplying Equation (29.31) by ryij and setting the result equal to zero yields

o=

ryi j U(ij)

= ryij Uij + Uk (2ryikUiu -

ryi j

xt

l)

+ UlUk(rykl Uuu -

- 2UklXik)ryil - 2UkUlmX~m)rykl - UiU jUkX~~ryij,

2xi:lryil) (29.32)

where we have used the wave equation, rykl ukl = O. Equation (29.32) must hold for all derivatives of U and powers thereof (treated as independent) modulo the

waveequation. Therefore, the coefficients of such "monomials" mustvanish. For example, since all the terms involving UkUl m are independent (even after substituting uxx + Uyy for Utt), we have to conclude that X~m) = 0 for all m, i.e., that X(i) are independent of u. Setting the coefficient of UkUl equal to zero and noting that Xi: l = ax~k) laxi = 0 yields Uuu

=0

=}

U(x, y, t, u)

= a(x, y, t)u + fJ(x, y, r).

(29.33)

Let us concentrate on the functions XCi). These are related via the term linear in Ukl. After inserting the wave equation in this term, we get UklXik)ryil = U12

(xf) -

+ U22 (X?)

X~l)) + U13 (X~3l - X~l)) - U23 (X~3) - X~2»)

+ U33 (X?)

- X~3l).

+ X~2»)

29.4 APPLICATION TO SOME KNOWN PDES

The

Uij

963

in this equation are all independent; so, we can set their coefficients equal

to zero: - X(2) - X(3) X (!) t 2 3

(29.34) The reader may verify that these relations imply that I. For example,

X~~2 = X~~2

=

Xj~2

= -

X~~3

= -

= -

X~?3

'-v--'

'-.-'

'-.-'

'-v--'

--

=xg~

=xj~~

=xji>r

::=Xg'3

= -

Xj~3

xj'11 = 0 for any i, j, k, and

x~;), = -

Xm = -X~~2'

=X~~l

=xWz

(29.35)

So, the first link ofthis chain is equal to its negative. Therefore, all the third derivatives in the chain of Equation (29.35) vanish. It follows that all Xli) 's are mixed polynomials of at most degree two. Writing the most general such polynomials for the three functions X(l), X(2), and X(3) and having them satisfy Equation (29.34) yields X(l)

+ a4X + a7Y + ast + as(x 2 + i + t 2) + 2a9xt + 2alQyt, 2) 2 a2 + asx - a6Y + a4t + a9(x - i + t + 2alQxy + 2asxt, 2) 2 a3 + a6X + asy + a7t + alQ( _x + i + t + 2U9xy + 2aSyt.

= a,

X(2) = X(3) =

(29.36) Setting the coefficient of Uk and ~ij Uij equal to zero and using Equation (29.33) gives 2a

x

= X(2) xx

+ X(2) _ yy

2a = X(!) t

tt

It follows that a(x, y, t)

X(l) xx

+ X(3) yy

X(2)

2a =

X(!)

fJtt - fJxx - fJyy = O.

ttl

yy'

y

x(3)

xx

_ X(2)

ttl

fJ is any solution of the wave equation, and

= au - ast - a9x - alQY.

By inserting the expressions found for Xli) and U in (29.30) and writing the result in the form Li aiVi, we discover that the generators of the symmetry group consist of the ten vector fields given in Table 29.3 as well as the vector fields

Uau,

vp = fJ(x, y, t)au

for fJ an arbitrary solution of the wave equation. The ten vector fields of Table 29.3 comprise the generators of the conformal group in three dimensions whose generalization to m dimensions was studied in Section 28.4.

964

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

Transformation

Infinitesimal generator VI V4 V6 V7

= a"

V2

= ax,

V3

Translation

= ay

=xa, + tax = -yax + xa y = ya, + tay

Vs = ta,

Rotation/Boost

+ xax + yay

Dilatation

= (t 2 + x 2 + y2)a, + 2xtax + 2ytay - tuau 2 2 V9 = 2xta, + (t + x - y2)ax + 2xya y - xua u 2 2 VlO = 2yta, + 2xyax + (t - x + y2)ay - yuau

VB

Inversions

Table29.3 The generators of the conformal groupfor]R3 , part of the symmetry groupof the wave equationin two dimensions.

29.5 Application to ODEs The theory of Lie gronps finds one of its most rewarding applications in the integration of ODEs. Lie's fundamental observation was that if one could come up with a sufficiently large group of symmetries of a system of ODEs, then one could integrate the system. In this section we outline the general technique of solving ODEs once we know their symmetries. The following proposition will be useful (see [Wam 83, p. 40]):

29.5.1. Proposition. LetM be an n-dimensional manifold andv E X(M).Assume that vip t= 0 for some P E M. Then there exists a local chart, i.e., local set of coordinate functions, (wI, ... , w") at P such that v = a/awl.

29.5.1 First-Order ODEs The most general first-order ODE can be written as

du dx

= F(x, u)

=}

Ll.(x, u, ux) es Ux - F(x, u)

= O.

(29.37)

29.5 APPLICATION TO ODES

965

A typical infinitesimal generator of the symmetry group of this equation is" v = X ax + u au, whose prolongation is pr(l) =

v+ u ex)_a ,

where

aux

u(x)

es U»

+ (Uu -

Xx)u x

-

Xuu~, (29.38)

as the reader may verify. The infinitesimal criterion for the one-parameter group of transformations G to be a symmetry group of Equation (29.37) is pr(1)v(t» = 0, or

au + (au _ ax) F_ axF2 = x aF + u aF. ax au ax au ax au

(29.39)

Any solution (X, U) of this equation generates a I-parameter group of transformations. The problem is that a systematic procedure for solving (29.39) is more difficult than solving the original equation. However, in most cases, one can guess a symmetry transformation (based on physical, or other, grounds), and that makes Lie's method worthwhile. Suppose we have found a symmetry group G with infinitesimal generator v that does not vanish at P E M C X x U. Based on Proposition 29.5.1, we can introduce new coordinates w = ~(x, u),

y = ~(x,u)

(29.40)

in a neighborhood of P such that v = ajaw, whose prolongation is also ajaw [see (29.38)]. This transforms the DE of (29.37) int05 2i.(y, w, w y ) = 0, and the infinitesimal criterion into pr(1)v(ii) =

aii aw

=

o.

It follows that ii is independent of w. The transformed DE is therefore ii (y, w y ) = 0, whose normal form, obtained by implicitly solving for dwjdy, is

-dw = H(y) dy

"*

w

==

l

Y

H(t)dt - w(a)

a

for some function H of y alone and some convenient point y = a. Substituting this expression of was a function of y in Equation (29.40) and eliminating y between the two equations yields u as a function of x. Thus our task is to find the change of variables (29.40). For this, we use v(w) = I and v(y) = 0, and express them in terms of x and u:

.

a~

a~

a~

a~ + u au

ax + uau

v(w) =v(~) = Xv(y) = v(~) = X ax

= I, = O.

(29.41)

4The reader is warned against the unfortunate coincidence of notation: X and U represent both the components of the infinitesimal generator andthe spacesof independent anddependent variables! 5Herewe arechoosing w to be the dependent variable. This choice is a freedom thatis alwaysavailable to us.

966

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

The second equation says that ry is an invariant of the group generated by v. We therefore nse the associated characteristic ODE [see (29.3) and (29.5)] to find y (or ry):

dx du = X(x, u) U(x, u)

(29.42)

To find W (or s), weintrodnce X(x, u, v) = v -Hx, u) and note that an eqnivalent relation containing the same information as the first eqnation in (29.41) is

x ax

ax

+U

ax au

+ aX

=0,

av

which has the characteristic ODE

dx X(x, u)

=

du U(x, u)

dv

(29.43)

= T'

for which we seek a solntion of the form v - !;(x, u) = c to read off Hx, u). The reader may wonder whether it is sane to go through so much trouble only to replace the original single ODE with two ODEs such as (29.42) and (29.43)1 The answer is that in practice, thelatter two DEs aremuch easier to solve than the original ODE.

29.5.2. Example. The homogeneous FODE dufdx = F(u/x) is invariant under the scaling transformation (x, u) t-+ (sx, su) whoseinfinitesimal generator is v = xa x + uau. Thefirst prolongation of thisvector is the sameas the vector itself (reader, verify!). To fiod the new coordinates w and y, first use Equation (29.42) with X(x, u) = x and U(x, u) = u:

. dx x

du u

u = C1 x

=> -

=>

u Y= x

(see Box 29.1.8).

Next, we note that (29.43) yields dx

du

-x = -u = dv =>

Inu

.

= v +luC2 => v = lu(u/C2)'

Substituting from the previousequation, we obtain v = lu(CjX/C2)

= lux

+ lu(CJ/C2)

=>

v -lux

= C => w = lux.

The chain rule gives duldx = (1 + yWy)/Wy. so that the DE becomes 1+ YWy = F(y) Wy

=>

dw = 1 , dy F(y)-y

which can be integrated to give w animplicit function of x.

= H(y) or lux = H(y) = H(u/x), which defines u as II

29.5 APPLICATION TO ODES

967

29.5.2 Higher-Order ODEs The same argument used in the first order ODEs can be used for higher-order ODEs to reduce their orders.

29.5.3. Proposition. Let ,;(x, u(n))

= ,;(x, u, UI, ... , un) = 0,

be an nth order ODE. Ifthis ODE has a one-parameter symmetry group, then there exist variables w = ~(x, u) and y = ~(x, u) such that A(X,

L3.

U

(n) ) __ /),. - (dw dnw) y, - , ... , - -_ 0 , dy dy"

i.e., in terms ofw and y, the ODE becomes of (n - I)st order in W y • Proof The proof is exactly the same as in the first-order case. The only difference is that one has to consider pr(n)y, where y = 8/8w. But Problem 29.7 shows that pr(n)v = v, as in the first-order case. 0 29.5.4. Example. Considera second-orderDE Ll.(u, ux, uxx) = 0, which does not de-

pendonx explicitly. The fact that aa/ax = osuggests w = x. So, we switch the dependent andindependent variables andwrite w = x, and y = u. Then, using the chainrule,we get

du

1

dx

= Wy

Snbstitutingin the originalDE, we obtain _

~(y, Wy, W y y)

w3 YY ) == A ( y, -1 , - = 0, wy

Wy

.

whichis of first order in Wy.

29.5.5. Example. The order of the SOLDE Uxx + p(x)ux + q(x)u = 0 can be reduced by notingthat the DE is invariant under thescalingtransformation (x, u) I-> (x, su), whose infinitesimal generator is v = uau • With this vector field, Equations (29.42) and (29.43) give dx

du

0=-;;-

dx = du = dv. o u

Forthe first equation to makesense, we haveto have

dx = 0

~

x = CI

~ Y=

x (by Box 29.1.8).

The secondequation in u gives v=lnu+c::::} v-lnu=c

=> w=lnu =>

u=ew .

968

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

Using thechain rule,we obtain ux

Riccati equation

w

dw w dw w rr-e =-e

dx

and

dy

By inserting this in theoriginal DE andwriting z = wy, we obtain

dz dy =

-Z

2

-

p(y)z - q(y),

whichis thewell-known first-order Riccati equation.

III

29.5.3 DEs with Multiparameter Symmetries We have seen that I-parameter symmetries reduce the order of an ODE by 1. It is natnral to suspect that an r-parameter symmetry will reduce the order by r. Although this suspicion is correct, it turns outthat in general, one cannotreconstrnct the solution of the original equation from those of the reduced (n - r )th-order equation. (See [Olve 86, pp. 148-158] for a thorough discussion of this problem.) However, the special, but important, case of second-order DEs is an exception. The deep reason behind this is the exceptional strnctnre of 2-dimensional Lie algebras given in Box 27.2.5. We cannot afford to go into details ofthe reasoning, but simply quote the following important theorem.

29.5.6.Theorem. Let ~(x, urn») = 0 be an nth-order ODE invariant under a 2-parameter group. Then there is an (n - 2)nd-order ODE li(y, w(n-2») = 0 with the property that the general solution to ~ can be found by integrating the general solution to li. In particular, a second-order ODE having a 2-parameter group of symmetries can be solved by integration. Let us analyze the case of a second-order ODE in some detail. By Box 27.2.5, the infinitesimal generators VI and V2 satisfy the Lie bracket relation

c=Oor1. We shall treat the abelian case (c = 0) and leave the nonabelian case for the reader.To begin with, we use sand t for the transformed variables, and at the end replace them with y and w. By Proposition 29.5.1, we can let VI = a/as. Then V2 can be expressed as the linear combination V2

a + P(s, t)-. a

= ",(s, t)as

The commutation relation

0= [a" ",a,

+ pa,j =

at

[VI, V2]

= 0 gives

a", -at as

ap

+ -a" as

29.5 APPLICATION TO ODES

969

showing that a and fJ are independent of s, We want to simplify V2 as mnch as possible without changing Vt. A transformation that accomplishes this is S = s + h(t) and T = T'tt), Then, by Equation (26.8) we obtain

a vi = Vj (S) as

a

as

aT

+ vr (T) aT = V2 = V2(S)~ + v2(T)~ =

, a

,a

ar a a + a:;;- aT = as' (a as + fJ as) ~ + (aaT + fJ aT) ~ as a as as as

at

as

as

at

er

( + fJh ) as + fJT ar

= a

If fJ

= 1/ fJ and h' = -a/ fJ to obtain a V2=-,

'" 0, we choose T'

at

where we have substituted s for S and t for T. If fJ change the notation from S to s and T to t to obtain

(29.44)

= 0, we choose a = T, and

a

(29.45)

V2 = t - .

as

The next step is to decide which coordinate is the independent variable, prolong the vector fields, and apply ittothe DE to find the infinitesimal criterion. For fJ '" 0, s and y t. Then the prolongation of Vj the choice is immaterial. So, let w and V2 will be the same as the vectors themselves, and with !;.(y, w, w y, wyy) es Wyy - F(y, w, wy), the infinitesimal criteria for invariance will be

=

=

a!;. aF = --, aw aw 2 a!;. aF 0= pre )V2(!;.) = V2(!;.) = = --. ay ay 2

0= pre )Vj(!;.) = vi (!;.) = -

It follows that in the (y, w) system, F will be a function of w y alone and the DE will be of the form

Wyy = F(w y)

=}

dw y dy

= F(w y)

=}

f

WY

~

F(z) .

~

=H(wy)

The last equation can be solved for w y in terms of y and the result integrated. For fJ = 0, choose w = t and y = s, Then Vj will not prolongate, and as the reader may verify, (2)

pr V2

= V2 -

a 2 a W- 3wywyy - aWyy Yaw y

a = way

2a a w y - - 3wywYY-aw' yy aw y

970

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

and the infinitesimal criteria for invariance will be .

aL'.

2

0= pr( )VI (L'.) = VI (L'.) = -

0=

pr(2)v2(L'.)

aF

= --,

ay ay er aF = -w -ay +w;aw y + 3wy _Wyy • '-..-'

=F

~O

It follows that in the (y, w) system, F will be a function of w and wyand satisfy the DE

aF = 3F , aw y

wy -

whosesolutionisoftheforrnF(w, w y) = w~F(w). The original DE now becomes 3Wyy = wyF(w),

for which we use the chain rule Wyy = WyaWy/aw to obtain

dw y 2dw = wyF(w)

1

=}

-

w = y

I

W -

F(z)dz

=}

dw 1 dy = - H(w)'

'-.----'

wH(w)

which can be integrated. Had we chosen w = s and y = t, F wonld have been a function of y and the DE wonld have reduced to Wyy = F(y), which conld be solved by two consecutive integrations. The nonabelian 2-dimensional Lie algebra can be analyzed similarly. The reader may verify that if fJ = 0, the vector fields can be chosen to be

a

VI =

as'

a as

(29.46)

V2 =S-,

leading to the ODE Wyy = wyF
a as

VI = - ,

a as

a at

V2=S-+t-,

i' 0, the vector fields can be chosen (29.47)

leading to the ODE Wyy = F(wy)/y. Both of these ODEs are integrable as in the abelian case.

29.6 Problems 29.1. Suppose that {Fi li'=1 are invariants ofthePDE (29.3). Show that any function f (FI, F2, ... , Fn ) is also an invariant of the PDE.

971

29.6 PROBLEMS

29.2. Find the function rotation of 80(2).

1 = IJ . I

when I(x)

= ax + b and IJ is the angle of

29.3. Use the resnlt of Problem 29.2 to find UI. Hint: Note that a = UI. 29.4. Transform the DE of Example 29.3.2 from Cartesian to polar coordinates to obtaindrjdlJ = r 3 . 29.5. Using the definition of total derivative, verify Eqnation (29.21). 29.6. Show that 80(2) is a symmetry group of the first-order DE L\.(x, u, UI)

= (u -

X)UI

+X +U = 0

and write the same DE in polar coordinates.

29.7. Show that the nth prolongation of the generator of the ith translation. the same as the original vector. 29.8. Find the first prolongation of the generator of scaling: xax

ai, is

+ uau •

29.9. Show that when the group acts only on the single dependent variable u, the prolongation of v = U au is given by P a pr(l)v=v+ L U j - , j=l aUj

au au where U· = - . +U·-. ) ax} } au

29.10. Show that the nth prolongation of v = X(x, u)ax ordinary DE of nth order is

+ U(x, u)au

for an

a

n

pr(n)V = v + "U[kl __ aU(k)' k=l

c:

where

u

(k) _

aku

-k - ax

and

29.11. Compute the second prolongation of the infinitesimal generators of the symmetry group of the heat equation. 29.12. Derive Equations (29.31) and (29.32). 29.13. Using Equation (29.34) show that Korteweg-de Vries

equation

XjZl = 0 for any i, j, k, and I.

29.14. The Korteweg-de Vries equation is u, + Ux xx + uUx = O. Using the technique employed in computing the symmetries of the heat and wave equations, show that the infinitesimal generators of symmetries of the Korteweg-de Vries

equation are

V3

=a V2 = 8" = tax + au,

V4

= xax

VI

X I

+ 31a, -

2uau ·

translation Galilean boost scaling

972

29. LIE GROUPS AND DIFFERENTIAL EQUATIONS

29.15. Suppose M(x, u) dx-s-Nix, u) du = Ohasa I-parameter symmetry group with generator v = xax + uau . Show that the function q(x, u) = I/(XM + UN) is an integrating factor. 29.16. Show that the second prolongation of v = Way (with y treated as independent variable) is a _ _ 3w a_. pr(2)v = v _ w 2_ ywyy _ aW yy Yaw y

29.17. Go through the case of f3 = 0 in the solution of the second order ODE and, choosing w = s and y = t, show that F will be a function of y alone and the original DE will reduce to W yy = F(y). 29.18. Show that in the case of the nonabelian 2-dimensional Lie algebra, (a) the vector fields can be chosen to be

a as

VI = - ,

a as

Vz = 8 -

if f3 = O. (b) Show that these vector fields lead to the ODE W yy = wyF(y). (c) If f3 'i- 0, show that the vector fields can be chosen to be VI

a as

=-,

V2

a +t-. a as at

= s-

(d) Finally, show that the latter vector fields lead to the ODE Wyy = F(wy)/y.

Additional Reading I. Blurnan, G. and Kumei, S. Symmetries andDifferential Equations, SpringerVerlag, 1989. A readable book on the subject of this chapter with many worked-out examples. 2. Olver, P. Application of Lie Groups to Differential Equations, SpringerVerlag, 1986. Our treatment of the subject follows closely that of Olver. This self-contained book, although formal, is very lucid in style with many historical notes. All the concepts introduced are clearly stated and many examples introduced to clarify them. 3. Stephani, H.Differential Equations: Their solutions using symmetries, Cambridge University Press, 1989. Written by a physicist, this little book treats solutions of ordinary differential equations in great detail with very little formalism.

30

-------

Calculus of Variations, Symmetries, and Conservation Laws

In this chapter we shall start with one of the oldest and most nseful branches of mathematical physics, the calcnlus of variations. After giving the fundamentals and some examples, we shall investigate the consequences of symmetries associated with variatioual problems. The chapter then ends with Noether's theorem, which counects such symmetries with their associated conservation laws. All vector spaces of relevance in this chapter will be assumed to be real.

30.1 The Calculus of Variations One of the main themes of calculus is the extremal problem: Giveu a function :J D -+ IR, find the points iu the domain D of 1 at which 1 attains a maximum or minimum. To locate such points, we find the zeros of the derivative of I. For multivariable functions, 1 : IRP :J n -+ IR, the notion of gradient generalizes that of the derivative. To find the jth component of the gradient VI, we calcnlate the difference 1'>.1 between the value of 1 at (xl, ... , x j +8, ... , x P) and its value at (xl, ... , xl , ... , x P), divide this difference by 8, and take the limit 8 -+ O. This is simply partial differentiation, and the jth component ofthe gradient is just the jth partial derivative of I.

1 : IR

30.1.1 Derivative for Hilbert Spaces To make contact with the subject of this chapter, let us reinterpret the notion of differentiation. The most useful interpretation is geometric. In fact, our first encounter with the derivative is geometrical: We are introduced to the concept through lines tangent to curves. In this language, the derivative of a function 1 : IR :J n -+ IR at Xo is a line (or function) '" : n :J no -+ IR passing through

974

30. CALCULUS OF VARIATIONS, SYMMETRIES, AND CONSERVATION LAWS

(xo, f(xo)) whose slope is defined to be the derivative of f atxo (see Figure 30,1): 1fJ(x) = f(xo)

+ f'(xo)(x -

xo).

The function 1fJ (x) describes a line, but it is not a linear function (in the vector-space sense of the word). The requirement of linearity is due to our desire for generalization ofdifferentiation to Hilbert spaces, on which linear maps are the most natural objects. Therefore, we consider the line parallel to 1fJ (x) that passes through the origin. Call this '" (x). Then (30.1)

"'(x) = !'(xo)x,

which is indeed a linear function, We identify "'(x) as the derivative of fat xo. This identification may appear strange at first but, as we shall see shortly, is the most convenient and useful. Of course, any identification requires a one-to-one correspondence between objects identified. It is clear that indeed there is a oneto-one correspondence between derivatives at points and linear functions with appropriate slopes. Equation (30.1) can be used to geometrize the definition of derivative. First consider

f ' (xo ) -- "'(x) - "'(xo)

x -xo

,

and

f'(xo) = lim f(x) - f(x o). x~xo

X -

xo

Next note that, contrary to f which is usually defined only for a subset of the real line, '" is defined for all real numbers rn;" and that "'(x - xo) = "'(x) - "'(xo) due to the linearity of "'. Thus, we have lim f(x) - f(xo) = "'(x) - "'(xo) = lim "'(x - xo) X--+XQ

x-

XQ

x-

XQ

X--+XQ

X -

xo '

or lim If(x) - f(xo) - "'(x - xo)1 = 0 Ix - Xo I

(30.2)

X-->Xo

where we have introduced absolute values in anticipation of its analogue-norm. Equation (30.2) is readily generalized to any complete normed vector space (Banach space), and in particnlar to any Hilbert space:

30.1.1. Definition. Let JfI and Jf2 be Hilbert spaces with norms II . III and II . 112, respectively. Let f : Jfl :J Q --* Jf2 be any map and Ixo) E Q. Suppose there is a linear map T E £.(Jfl, J(2) with the property that lim IIf(lx}) - f(lxo)) - T(lx) - Ixo))112 = 0 IIx-xo II 1-->0 IIx - xolll differentiability ofa function on aHilbert space ata point

for Ix)

E Q.

Then, we say that f is differentiable at Ixo}, and we define the derivative of f at Ixo} to be Df(xo) ea T. If f is differentiable at each Ix) E Q, the map given by is called the derivative of f.

Df(lx)) = Df(x)

30.1 THE CALCULUS OFVARIATIONS

975

!

!(xo) - - - - - - -

~---Q-----iIo-

Figure 30.1 Thederivative at (xQ. f (xo» as a linearfunction passingthrough theorigin with a slope j'(xo). Thefunction f is assumed to be defined on a subset {2 of the real line. 00 restricts the x's to be close to XQ to prevent the function from misbehaving (blowing up),andto make sure that thelimitin thedefinition of derivative makes sense.

The reader may verify that if the derivative exists, it is unique. 30.1.2. Example. Let 1f1 = lll.n and 1f2 = lll.m and f : lll.n :J n -+ lll.m. Then for [z} E n, Df(x) is a linear map, which can be represented by a matrix in the standard bases of ~n andIRm. Tofindthis matrix, we needto let Df (x) actonthe jth standard basisof R" • i.e.•we need to evaluate Df(x) lei)' Thissuggests taking Iy) = Ix) + h lei) (with h -+ 0) as thevector appearing in thedefinition of derivative at Ix}. Then IIf(ly» - f(lx» - Df(xHly)

-lx»)112

lIy -xIII IIf(x l , ... ,xi

+ h, ... , x n) -

f(x l, ... , xi, ... , x n) - hDf(x) lei)1I 2

Ihl approaches zero as h -+ 0, so that the ith component of the ratio also goes to zero. But the ith component of Df(x) lei) is simply a~, the ijth component of the matrix of Df(x).

Therefore, .

Iim

h....O

x " If i (x l , ... ,Xi+h , ... ,Xn) - hfill

which means that a~

.. ,X i

, ... ,Xn) -

I I

= a/ jax i .

The result of the example above can be stated as follows:

hil aj

=0, III

976

30. CALCULUS OF VARIATIONS, SYMMETRIES, ANO CONSERVATION LAWS

30.1.3. Box. For f : R" :J n --+ R"', the matrix of Df (x) in the standard basis ofR" and R'" is the Jacobian matrix of f.

differential and gradient of f at Ix)

The case of:H2 = R deserves special attention. Let:H be a Hilbert space. Then Df(x) E £(:H, R) =:H* is denoted by df(x) and renamed the differential of f at [x). Furtbennore, through the inner product, one can identify d f : n --+ :H* with another map defined as follows:

30.1.4. Definition. Let:H be a Hilbert space and f : :H :J

n --+ R. The gradienl

V f of f is the map V f : n --+ :H defined by (V f(x)1 a) '" (df(x), a)

V Ix) En, 10) E:H

where (, ) is the pairing (, ) : :H* x :H --+ R of:H and its dual. Note that although f is not an element of:H*, df(x) is, for all points Ix} E at which the differential is defined.

30.1.5. Example. Consider the function I : X

-+

n

rn:. givenby I(lx» = IIx 11 2 Since

lIy - xll 2 = lIyll2 -lIxll 2 - 2 (xl y - x)

and since the derivative is unique, the reader may check that d/(x) 10) = 2 (x] a), or V 1(lx)) = 2Ix). III Derivatives could be defined in terms of directions as well:

n --+ :H2 be has a derivative in the direction 10) E :HI at

30.1.6, Definition. Let:HI and:H2 be Hilbert spaces. Let f : :HI :J directional derivative

any map and Ix) E [x) if

n. We say that f

exists. We call this element of:H2 the directional derivative of fin Ihe direction 10) E :HI at [x). The reader may verify that if f is differentiable at [z) (in the context of Definition 30.1.1), then the directional derivative of f in any direction 10) exists at Ix} and is given by

!:.f(lx) + I lanl = dt ,=0

Df(x) 10).

(30.3)

30.1 THE CALCULUS OF VARIATIONS

30.1.2

971

Functional Derivative

We now specialize to the Hilbert space of square-integrable functions .c,2(Q) for some open subset Q of some JRm. We need to change our notation somewbat. Let us agree to denote the elements of .c,2(Q) by f, u, etc. Real-valued functions on .c,2(Q) will be denoted by L, H, etc. The m-tuples will be denoted by boldface

lowercase letters. Tosummarize, x, Y E JRm, f, u E .c,2(Q)

UI u) =

functional derivative orvariational derivative

1

"*

f, u : JRm ::>

Q --+ JR,

L, H : .c,2(Q) --+ JR.

f (x)u (x) dmx,

Furthermore, the evaluation of L at u is denoted by L[u]. When dealing with the space of functions, the gradient of Definition 30.1.4 is called a functional derivative or variational derivative and denoted by 8L/8u.

So

~L l f) = Jgr 8L(x)f(x) dmx = ( uU 8u

dd L[u+tfJl ' t 1=0

(30.4)

where we have used Equation (30.3). Note that by definition, 8L/8u is an element of the Hilbert space .c,2(Q); so, the integral of (30.4) makes sense. Equation (30.4) is frequently used to compute functional derivatives. An immediate consequence of Equation (30.4) is the following important result. 30.1.7. Proposition. Let L : .c,2(Q) --+ JRforsome Q C JRm. IfL has an extremum

at u, then 8L -=0.

8u

Proof If Lhas an extremum at u, then the RHS of (30.4) vanishes for any function f, in particular, for any orthonormal basis vector lei). Completeness of a basis now implies that the directional derivative must vanish (see Proposition 5.1.9). 0

evaluation function

Just as in the case of partial derivatives, where some simple relations such as derivative of powers and prodncts can be used to differentiate more complicated expressions, there are some primitive formulas involving functional derivatives that are useful in computing other more complicated expressions. Consider the evalnation fnnction given by

Ey[fJ = fey).

Using Equation (30.4), we can easily compute the functional derivative of Ey:

r

)0.

8E;[U] (x)f(x) dmx = dd Ey[u + tfJl = dd luCY) + tf(y)} uU t t=O t 8E [u] = fey) - y-(x) = 8(x - y).

"*

8u

I

1=0

(30.5)

978

30. CALCULUS OF VARIATIONS, SYMMETRIES, AND CONSERVATION LAWS

It is instructive to compare (30.5) with the similar formula in multivariable calculus, where real-valued functions f take a vector x and give a real number. The analogue of the evaluation function is Ei, which takes a vector x and gives the real number xi, the ith component of x. Using the definition of partial derivative, one readily shows that 8E;/8x i = 8ij, which is (somewhat less precisely) written as 8x i j8x i = 8ij. The same sort of imprecision is used to rewrite Equation (30.5) as 8u(y)

8u y

8u(x)

8ux

- - == -

= 8(x - y),

(30.6)

where we have turned the arguments into indices to make the analogy with the

discrete case'even stronger. Another useful formula concerns derivatives of square-integrable functions. Let Ey,i denote the evaluation of the derivative of functions with respect to the ith coordinate: given by

Ey,i (f) =

8;j(y).

Then a similar argument as above will show that

8Ey i

- - ' (x) = -8i8(X - y), 8u

or

88iU (y ) 8u(x) = -8i 8(X - y),

and in general,

88il...i,U(y) = (-1)k8' . 8(x _ ). 8u(x) 'I .. ·" y

(30.7)

Equation (30.7) holds only if the function f, the so-called test function, vanishes on 8rl, the boundary of the region of integration. If it does not, then there will be a "surface term" that will complicate matters considerably. Fortunately, in most applications this surface term is required to vanish. So, let us adhere to the

convention that

30.1.8. Box. All test functions f(x) appearing in the integral of Equation (30.4) are assumed to vanish at the boundary ofrl.

For applications, we need to generalize the concept of functions on Hilbert spaces. First, it is necessary to consider maps from a Hilbert space to JR". For simplicity, we confine ourselves to the Hilbert space .c,2(rl). Such a map H .c,2(rl) --> D C JR", for some subset D of JR", can be written in components where

Hi: .c,2(rl) --> JR,

i = I, ... , n.

979

30.1 THE CALCULUS OF VARIATIONS

Next, we consider an ordinary multivariable function L : IR!.n ::J D --> IR!., and use it to construct a new function on ,(,2(\1), the composite of Land H: La H[u] = L (HI[U], ... , Hn[u]).

La H : ,(,2(\1) --> IR!.,

Then the functional derivative of L a H can be obtained using the chain rule and noting that the derivative of L is the common partial derivative. It follows that

~L a H[u] -::-"-"(x) = ~U

Is

}

n

~Hi

- L (HI[U], ... , Hn[u]) (x) = I)iL-(X), ~U i=1 . ~U

(30.8)

where aiL is the par1ial derivative of L with respect to its ith argument.

30.1.9. Example. Let L : (a, b) x Ill. x Ill. -+ R, be a function of three variablesthe first one of which is definedfor the real interval (a, b). Let Hi : £}(a, b) -+ Ill.,i = 1,2,3,be definedby H2[U] = Ex[u] = u(x), where Ex is theevaluation function andei evaluates thederivative. Itfollowsthat L oH[u] = L(x, u(x), u'(x)). Then, noting that HI[U] is independent ofu, we have 8L a H[u] 8HI[u] 8u (y) = oIL~(y)

8E~[u]

8Ex[u]

+ 02L~(y) + 03L~(y)

= 0 + BzL8(y - x) - 03H' (y - x) = 02H(X - y)

+ 03H'(X -

y).

Thisis normally written as 8L(x, u(x), u'(x)) ()

y =

Bu

az'(x - y ) + az ,,(x - y ), au au'

-0

-0

(30.9)

which is the unintegrated version of the classical Euler-Lagrange equation for a single particle, to which we shallreturn shortly. ..

A generalization of the example above toms L into a function on \1 x IR!. x IR!.m with \1 C IR!.m, so that L (xl, ... , x m, u(x), aIU(X), ... , amu(x)) E IR!.,

The functions {Hi Hi[U]

17:t

l

are defined as

== xi

for i

Hi[U] sa Ex[u] = u(x) Hi[U] sa Ex,i[U]

= aiU(X)

= 1,2, ... ,m,

+ I, m + 2, ... , 2m + I,

for i = m for i =

and lead to the equation ~L a H[u]

_-::--=--.:c(y) = am+IL~(x - y) Bu

+

2m+1

L

aiLai~(X - y),

(30.10)

i=m+2

which is the unintegrated version of the classical Euler-Lagrange equation for a field in m dimensions.

980

30. CALCULUS OF VARIATIONS, SYMMETRIES, AND CONSERVATION LAWS

30.1.3 Variational Problems The fundamental theme of the calculus of variations is to find functions that extremize an integral and are fixed on the boundary of the integration region. A prime example is the determination of the equation of the curve of minimum length in the xy-plane passing through two points (XI, Yl) and (X2, Y2). Such a curve, written as Y = u(x), would minimize the integral Int[u] sa

l

x2

JI + [u'(x)]2dx,

U(XI)

= YI,

U(X2)

= Y2·

(30.11)

xr

Note that int takes a function and gives a real number, i.e.-if we restrict our functions to square-integrable ones-intbelongs to ,c2(XI, X2). This is how contact is established between the calculus of variations and what we have studied so far in this chapter. To be as general as possible, we allow the integral to contain derivatives up to the nth order. Then, using the notation of the previous chapter, we consider functions L on MCn) cOx U Cn), where we have replaced X with 0, so that M = jRP ::J 0 x U C jRq. nth-order variational problem; Lagrangian; functional

30.1.10. Definition. By an nth-order variational problem we mean finding the extremum of the real-valued function L : ,c2(0) -+ jR given by L[u] sa

in

L(x, uCn») dPx,

(30.12)

where 0 is a subset ofjRP x jRqp(n), L is a real-valued function on 0, and pCn) = (p + n)!j(n!p!). In this context the function L is called the Lagrangian of the problem, and L is called a functional. I The solution to the variational problem is given by Proposition 30.1.7, moving the functional derivative inside the integral, and a straightforward (but tedious!) generalization of Equation (30.10) to include derivatives of order higher than one. Due to the presence of the integral, the Dirac delta function and all its derivatives will be integrated out. Before stating the solution of the variational problem, let us introduce a convenient operator. Euler operator

30.1.11. Definition. For I lEa=L(-DlJ J

:s: a :s: q, the ctth Euler operator is

a

(30.13)

auJa'

where for J = (ii, ... , AJ, (-D)J sa (-I)kDJ = (-Dh)(-Dj,)'" (-DA),

and the sum extends over all multi-indices J 1Do not confuse this functional with the linear functional of Chapter 1.

= (iI, ... , ik), including J = O.

30.1 THE CALCULUS OF VARIATIONS

981

The negative signs are introduced because of the integration by parts involved in the evaluation of the derivatives of the delta function. Although the sum in Equation (30.13) extends over all multi-indices, only a finite nmnber of terms in the sum will be nonzero, because any function on which the Euler operator acts depends on a finite number of derivatives. 30.1.12. Theorem. If u is an extremal of the variational problem (30.12), then it

Euler-Lagrange equations

must be a solution ofthe Euler-Lagrange equations ex = 1, ... ,q.

Leonhard Euler (1707-1783) was Switzerland's foremost scientist and one of the three greatest mathematicians of modem times (Gauss and Riemann being the other two). He was perhaps the most prolific author of all time in any field.

Fram 1727 to 1783 his writings poured out in a seemingly endless flood, constantly adding knowledge to every known

branch of pure and applied mathematics, and also to many that were not known until he created them. He averaged about 800 printed pages a year throughout his Iong life, and yet he almost always had something worthwhile to say. The publication of his complete works was started in 1911, and the end is not in sight. This edition was planned to include 887 titles in 72 volumes, but since that time extensive new deposits of previously unknown manuscripts have been unearthed, and it is now estimated that more than 100 large volumes will be required for completion of the project. Euler evidently wrote mathematics with the ease and fluency of a skilled speaker discoursing on subjects with which he is intimately familiar. His writings are models of relaxed clarity. He never condensed, and he reveled in the rich abundance of his ideas and the vast scope of his interests. The French physicist Arago, in speaking of Euler's incomparable mathematical facility, remarked that "He calculated without apparent effort, as men breathe, or as eagles sustain themselves in the wind." He suffered total blindness

during the last 17 years of his life, but with the aid of his powerful memory and fertile imagination, and with assistants to write his books and scientific papers from dictation, he actually increased his already prodigious output of work. Euler was a native of Basel and a student of Johann Bernoulli at the University, but he soon outstripped his teacher. He was also a man of broad culture, well versed in the classica11anguages and literatures (he knew the Aeueid by heart), many madera languages, physiology, medicine, botany, geography, and the entire body of physical science as it was known in his time. His personal life was as placid and uneventful as is possible for a man

with 13 childreu.

.

Though he was not himself a teacher, Euler has had a deeper influence on the teaching of mathematics than any other person. This came about chiefly through his three great treatises: Introductio in Analysin Infinitorum (1748); Institutiones Calculi Differentialis (1755); and lnstitutiones Calculi Integralis (1768-1794). There is considerable truth in the old saying

982

30. CALCULUS OF VARIATIONS, SYMMETRIES, AND CONSERVATION LAWS

that all elementary and advanced calculus textbooks since 1748 are essentially copies of Euler or copies of copies of Euler. These works summed up and codified the discoveries of his predecessors, and are full of Euler's own ideas. He extended and perfected plane and solid analytic geometry, introduced the analytic approach to trigonometry, and was responsible for the modem treatment of the functions In x and e", He created a consistent theory of logarithms of negative and imaginary numbers. and discovered that In x has an infinite number of values. It was through his work that the symbols e. 1C. and i = J=T became common currency for all mathematicians, and it was he who linked them together in the astonishing relation e in = -1. Among his other contributions to standard mathematical notation were sin x, cosx, the use of f(x) for an unspecified function, and the use of L for summation. His work in all departments of analysis strongly influenced the further development of this subject through the next two centuries. He contributed many important ideas to differential equations, including substantial parts of the theory of second-order linear equations and the method of solution by 'power series. He gave the first systematic discussion of the calculus of variations, which he founded on his basic differential equation for a minimizing curve. He discovered the integral defining the gamma function and developed many of its applications and special properties. He also worked with Fourier series, encountered the Bessel functions in his study of the vibrations of a stretched circular membrane, and applied Laplace transforms to solve differential equations-all before Fourier, Bessel, and Laplace werebom. E. T. Bell, the well-known historian of mathematics, observed that "One of the most remarkable features of Euler's universal genius was its equal strength in both of the main currents of mathematics, the continuous and the discrete." In the realm of the discrete, he was one of the originators of number theory and made many far-reaching contributions to this subject throughout his life. In addition, the origins of topology-cone of the dominant forces in modem mathematics-lie in his solution of the Konigsberg bridge problem and his formula V - E + F = 2 connecting the numbers of vertices, edges, and faces of a simple polyhedron. The distinction between pure and applied mathematics did not exist in Euler's day, and for him the entire physical universe was a convenient object whose diverse phenomena offered scope for his methods of analysis. The foundations of classical mechanics had been laid down by Newton,but Euler was the principal architect. In his treatise of 1736 he was the first to explicitly introduce the concept of a mass-point, or particle, and he was also the first to study the acceleration of a particle moving along any curve and to use the notion of a vector in connection with velocity and acceleration. His continued successes in mathematical physics were so numerous, and his influence was so pervasive, that most of his discoveries are not credited to him at all and are taken for granted in the physics community as part of the natural order of things. However, we do have Euler's angles for the rotation of a rigid body, and the all-important Euler-Lagrange equation of variational dynamics. Euler was the Shakespeare of mathematics-universal, richly detailed, and inexhaustible.

30.1 THE CALCULUS OF VARIATIONS

983

= q = 1, the Euler operator becomes a ~ ·a a a + D za - ... ,. IE = - + L,,(-Dx)J- = - - D For the special case of p

au

auj

j=l

xaux

au

x-[Ju xx

where Dx is the total derivative with respect to x, aod U j is the jth derivative of with respect to x; aod the Euler-Lagrange equation for the variational problem

U

becomes

aL

IE(L) = au

n

.'

aL

+ L(-l)l D i - = o.

(30.14)

aUj

j=l

Since L carries derivatives up to the n-th order aod each Dx carries one derivative, we conclude that Equation (30.14) is a 2n-th order ODE.

30.1.13. Example. Thevariational problem of Equation (30.11) hasa Lagraugiau L(u, u Cn)

= L(u, u(1)

=

J1+ ui,

which is a function of the first derivative only. So, the Euler-Lagrange equation takes the

form

et.

O=-Dx a ux =-Dx or uxx

=

0, so that u

x) d ( x ) (u Jl+ui =-dx Jl+ui =

= f(x)

U

= cj x

+ cz. The solution to the variational problem is a

straight line passing through thetwopoints (xj , YI) aud (xz, yz). The first variafion is not sufficientfor a full knowledge ofthe nature ofthe extremum!

uxx

III

The variational problem is a problem involvingamy the first functionalderivative, or the first variation. We know from calculus that the first derivativeby itself caooot detennioe the nature of the extremum. To test whether the point in question is maximum or mioimum, we need all the second derivatives (see Exarople 4.7.4). One uses these derivativesto expaod the functional 10a Taylor series up to the second order. The sign of the second order contribution determines whether the functional is maximum or minimum at the extremal point, 10 aoalogy with Exarople 4.7.4, we expaod L[u] about f up to the second-order derivative: L[u] = L[f]

+

+(



I

dPy , 8(L) (u(y) - f(y)) aU Y u=f

zL

I

1 ({ 8 2:}0, dPy }o, dPy' 8u(y)8u(y') u=f (u(y) - f(Y))(u(y') - f(Y'))·

984

30. CALCULUS OF VARIATIONS, SYMMETRIES, AND CONSERVATION LAWS

The integrals have replaced the sums of the discrete case of Taylor expansion of the multivariable functions. Since we are interested in comparing u with the f that extremizes the functional, the second term vanishes and we get L[u] = L[f]

1 + 2:11

2L

8 n dPy n dPy' 8u(y)8u(y')

Iu~f

. [(u(y) -I(y»(u(y') - f(y'»]·

(30.15)

Joseph Louis Lagrauge (1736-1813) was born Giuseppe Luigi Lagrangia but adopted the French version of his name. He was the eldest of eleven children. most of whom did not reach adulthood. His father destined him for the Iaw-a profession that one of his brothers later pursued-and Lagrange offered no objections. But having begun the study of physics and geometry, he quickly became aware of his talents and henceforth devoted himself to the exact sciences. Attracted first by geometry, at the age of seventeen he turned to analysis, then a rapidly developing field. In 1755, in a letter to the geometer Giulio daFagnano, Lagrange speaks of one of Euler's papers published at Lausanne and Geneva in 1744. The same letter shows that as early as the end of 1754 Lagrange had found interesting results in this area, which was to become the calculus of variations (a term coined by Euler in 1766). In the same year, Lagrange sent Euler a summary, written in Latin, of the purely analytical method that he used for this type of problem. Euler replied to Lagrange that he was very interested in the technique. Lagrange's merit was likewise recognized in Turin; and he was named, by a royal decree, professor at the Royal Artillery School with an annual salary of 250 crowns-a sum never increased in all the years he remained in his native country. Many years later, in a letter to d'Alemberl, Lagrange confirmed that this method of maxima and minima was the first fruit of his studies-he was only nineteen when he devised it-and that he regarded it as his best work in mathematics. In 1756, in a letter to Euler that has been lost, Lagrange, applying the calculus of variations to mechanics, generalized Euler's earlier work on the trajectory described by a material point subject to the influence of central forces to an arbitrary system of bodies, and derived from it a procedure for solving all the problems of dynamics. In 1757 some young Turin scientists, among them Lagrange, founded a scientific society that was the origin of the Royal Academy of Sciences of Turin. One of the main goals of this society was the publication of a miscellany in French and Latin, Miscellanea Taurinensia ou Melanges de Turin, to which Lagrange contributed fundamentally. These contributions included works on the calculus of variations, probability, vibrating strings, and the principle of least action. To enter a competition for a prize, in 1763 Lagrange sent to the Paris Academy of Sciences a memoir in which he provided a satisfactory explanation of the translational motion ofthe moon. In the meantime, the Marquis Caraccioli, ambassador from the kingdom of Naples to the court of Turin, was transferred by his government to London. He took along the young Lagrange, who until then seems never to have left the immediate vicinity of Turin. Lagrange was warmly received in Paris, where he had been preceded by his memoir on lunar

30.1 THE CALCULUS OF VARIATIONS

985

libratian.Hemayperhaps havebeentreatedtoowell in theParisscientific community, where austerity was not a leading virtue. Being of a delicate constitution, Lagrange fell ill and had to interrupt bis trip. In the spring of 1765 Lagrange returned to Tnrin by way of Geneva. In the autumn of 1765 d'Alembert, who was on excellent tenus with Frederick IT of Prussia, and familiar with Lagrange's work through Melanges de Turin, suggested to Lagrange that he accept the vacant position in Berlin created by Euler's departure for St. Petersburg. 11seems quite likely that Lagrange would gladly have remained in Tnrin bad the court of Turin been willing-to improve his material and scientific situation. On 26 April,

d' Alembert transmitted to Lagrange the very precise and advantageous propositions of the king of Prussia. Lagrange accepted the proposals of the Prussian king and, not without difficulties, obtained his leave through the intercession of Frederick II with the king of Sardinia. Eleven months after his arrival in Berlin, Lagrange married his cousin Vittoria Conti wbo died in 1783 after a long illness. With the death of Frederick II in August 1786 he also lost his strongest support in Berlin. Advised of the situation, the princes of Italy zealously competed in attracting him to their courts. In the meantime the French government decided to bring Lagrange to Paris through an advantageous offer. Of all the candidates, Paris was victorious. Lagrange left Berlin on 18 May 1787 to become pensionnaire veteran of the Paris Academy ofSciences, of which he had been a foreign associate member since 1772. Wannly welcomed in Paris, he experienced a certain lassitude and did not immediately resume his research. Yet he astonished those around him by his extensive knowledge of metaphysics, history, religion, linguistics, medicine, and botany. In 1792 Lagrange married the daughter of his colleague at the Academy, the astronomer Pierre Charles Le Monnier. This was a troubled period, about a year after the flight of the king and his arrest at Varennes. Nevertheless, on 3 June the royal family signed the marriage contract "as a sign of its agreement to the union." Lagrange had no children from this second marriage, which, like the first, was a happy one. When the academy was suppressed in 1793, many noted scientists, including Lavoisier, Laplace, and Coulomb were purged from its membership; but Lagrange remained as its chairman. For the next ten years, Lagrange survived the turmoil of the aftermath of the French Revolution, but by March of 1813, he became seriously ill. He died on the morning of 11 April 1813, and three days later his body was earned to the Pantheon. The funeral oration was given by Laplace in the name of the. Senate.

Astraight line segment is indeed lhe shortest distance between two points.

30.1.14. Example. Let us apply Equation (30.15)to the extremal function of Example 30.1.13 to see if the line is truly the shortest distance between two points. The firstfunctional derivative, obtained using Equation (30.9), is simply IE(L):

8L

- - =IE(L) = 8u(y)

Uyy

To find the second variational derivative, we use the basic relations (30.6), (30.7), and the cbain rule (30.10):

2L 8

I =-

8u(y')8u(y) u=f

8 8u(y')

[

Uyy

(1 + u~)3/2

]

Iu=f

986

30. CALCULUS OF VARIATIONS, SYMMETRIES, ANO CONSERVATION LAWS

= -{(I

=

+ u 2)-3/2 y

B"(y-y') ( I + u 2)3/2

I

y

because Uyy

= 0 and Uy

L[u] = L[f] -

= L[f] -

= C1 when u 1

2(1 + c2l)

Buyy _ u J(1 Bu(y') YY2

=

u=f

= f.

+ u 2)-5/2 2u ~ II YBu(y')

y

u~f

B"(y-y') (I + c21 )3/2 '

Inserting this in Equation (30.15), we obtain

3/21X2dY lX2dy'B" (y - y')(u(y) - f(y))(u(y') - f(y')) X[

XI

2 I lX2 d 23/2 dy(u(y) - f(y))-d2 (u(y) - f(y))· 2(1+c l) X[ Y

Thelastintegral canbe integrated by parts, withtheresult

l

X2

d (u(y) - f(y))"d(u(y) - f(y)) Y

I

-

Xl,

x2

Xl

d dy [ "d(u(y) - f(y)) Y

]2

=Obecauseu(Xj) =f(Xj), i = 1,2

Therefore, L[u]

= L[f] +

2(1

I2 3/2 lX2 dy [ -dd (u(y) -

+ c l)

~x..:[

y~

f(y))

]2

~

always positive

It follows that L[f] < L[u], i.e., that f indeed gives the shortest distaoce.

III

30.1.15. Example. Io the specialtheoryof relativity,the element of the invatiaot"length," or proper time, is given by ,j dt 2 - dx 2. Thus, the total proper time between two events (tl, Xl) aod (t2, X2) is given by L[x]

=1

'2

jl-xTdt,

dx Xt

II

ea dt '

The extremum of this variational problem is exactly the sameas in the previous example, theonly difference being a sign. In fact. thereader mayverifythat BL[x] = IE (L) = X" , Bx(s) (I - x'f)3/2

connection between variational problem and the twin paradox

and therefore, x = f(t) = qt + C2 extremizes the functional. The second variational derivative canbe obtained asbefore. It is leftforthereader to showthat in thecaseathand, L[f] > L[x], i.e., that f gives the longest proper time. Since the function f(t) = Cjt + C2 corresponds to aninertial (unaccelerated) observer, we conclude that

30.1.16. Box. Accelerated observers measure a shorter proper time between any two events than inertial observers.

This is the content of the famous twin paradox, in which the twin who goes to a distant galaxy andcomes back (therefore being accelerated) will return younger than her III (noaccelerated) twin.

30.1 THE CALCULUS OF VARIATIONS

987

30.1.4 Divergence and Null Lagrangians The variational problem integrates a Lagrangian over a region Q of JRP. If the Lagrangian happens to be the divergence of a function that vanishes at the boundary of Q, the variational problem becomes trivial, because all functions will extrentize the functional. We now study such Lagrangians in more detail. total divergence

30.1.17. Definition. Let (Fi : M(n) --+ JR}r~1 be functions an (Fl, ... , Fp ) . The total divergence ofF is defined to bt?

u»,

and F =

P

D·F", LDjFj, j=l

where D j is the total derivative with respect to xi. Now suppose that the Lagrangian L(x, u(n» can be written as the divergence of some p-tuple F. Then by the divergence theorem, L[uJ= {L(x,u(n»dPx= {D.FdPx= (

In

null Lagrangians

In

Jan

F·da

for any u = f(x) and any domain Q. It follows that L[J] depends on the behavior of f only at the boundary. Since in a typical problem no variation takes place at the boundary, all functions that satisfy the boundary conditions will be solutions of the variational problem, i.e., they satisfy the Euler-Lagrange equation. Lagrangians that satisfy the Euler-Lagrange equation for all u and x are called null Lagrangians. It turns out that null Lagrangians are the only such solutions of the Euler-Lagrange equation (for a proof, see [Olve 86, pp. 252-253]). 30.1.18. Theorem. Afunction L(x, u(n» satisfiesE (L) es Ofor all x andu if and only if L = D· Fforsome p-tuple offunctionsF = (Fl, ... , F p ) ofx, u, and the derivatives of u.

In preparation for the investigation of symmetries of the variational problems, we look into the effect of a change of variables on the variational problem and the Euler operator. This is important, because the variational problem should be independent of the variables chosen. Let

x=

W(x, u),

it = cf>(x, u)

(30.16)

be any change of variables. Then by prolongation, we also have it(n) = cf>(nl(x, u(n» for the derivatives. Substituting u = f(x) and all its prolongations in terms of the new variables, the functional L[J] =

l

L(x, pr(n) f(x» dPx

2The reader need not be concerned about lack of consistency in the location of indices (upper vs. lower), because we are dealing with indexed objects, such as Fi' which are not tensors!

9BB

30. CALCULUS OF VARIATIONS, SYMMETRIES, ANO CONSERVATION LAWS

will be transformed into

where the transformed domain, defined by

Q = Ii

= lj/(x, f(x» I x E Qj,

will depend not only on the original domain Q, bnt also on the fnnction f. The new Lagrangian is then related to the old one by the change of variables formnla for multiple integrals:

L(x, pr(n) f(x)) = L(i,

pr(n) !(i))detJ(x,

prel)f(x)) ,

(30.17)

where J is the Jacobian matrix of the change of variables induced by the fnnction



Starting with Equations (30.16) and (30.17), one can obtain the transformation formula for the Euler operator stated below. The details can be fonnd in [Olve 86, pp. 254-255j.

30.1.19. Theorem. Let L(x, u(n) and L(i, a(n) be two Lagrangians related by the change ofvariable formulas (30.16) and (30.17). Then ex = 1, ... ,q

where IE p is the Euler operator associated with the new variables, and

30.2

Symmetry Groups of Variational Problems

In the theory offields, as well as in mechanics, condensed matter theory, and statistical mechanics, the starting point is usually a Lagrangian. The variational problem of this Lagrangian gives the classical equations of motion, and its symmetries lead

to the important conservation laws. variational symmetry group

30.2.1. Definition. A local group of transformations G acting on M C Qo x U is a variational symmetry group ofthe junctional (30.18)

30.2 SYMMETRY GROUPS OF VARIATIONAL PROBLEMS

989

ifwhenever (the closure of) Qlies in Qo, f is afunction over Q whose graph is in M, and g EGis such that j = g . f is a single-valued function defined over fl, then

In "Symmetry ofthe Lagrangian" is really thesymmetry group ofthe variational problem!

l

L(Ji, pr(n) j(Ji)) dPJi =

L(x, pr(n) f(x)) dPx.

(30.19)

In the physics community, the symmetry group of the variational problem is (somewhat erroneously) called the symmetry of the Lagrangian. Note that if we had used L in the LHS of Equation (30.19), we would have obtained an identity valid for all lagrangians because of Equation (30.17) and the formula for the change in the volume element of integration. Ouly symmetric Lagrangians will satisfy Equation (30.19). As we have experienced so far, the action of a group can be very complicated and very noulinear. On the other hand, the infinitesimal action simplifies the problem considerably. Fortunately, we have (see [Olve 86, pp. 257-258] for a proof).

30.2.2. Theorem. A local group oftransformations G acting on M C Qo x U is a variational symmetry group ofthe functional (30.18) ifand only if pr(n)v(L)

+ LD . X = 0

for all (x, u(n») E

M(n)

a

P V=

and every infinitesimal generator

LXi(x,u)-.

ax'

;=1

(30.20)

a

q

+ LUa(x,u)-----;; a=l

au

ofG, where X sa (Xl, ... , X P ) . 30.2.3. Example. Considerthe case of p = I = q, and assumethat the Lagrangian is independent of x butdepends on U E C 2 (a , b) andits first derivative. Thenthe variational problem takes the form

1 b

L[u] =

1 b

L(u(l»)dx es

L(u,ux)dx.

Since derivatives are independent of translations, we expect translations to be part of the symmetry group of thisvariational problem. Let us verifythis. Theinfinitesimal generator of translation is ax, which is its own prolongation. Therefore, with X = 1 and U = 0, it

follows that pr(l)v(L) + LD· X

= "xL +

LDxX

= 0 +0 = O.

III

30.2.4. Example. As a less trivial case, consider the proper time of Example 30.1.15. Lorentz transformations generated by3 v

= u8x + x8u are symmetries of thatvariational

310order to avoid Confusion in applying formula (30.20), we use x (instead of t) as the independent variable andu (instead of x) as thedependent variable.

990

30. CALCULUS OF VARIATIONS, SYMMETRIES, ANO CONSERVATION LAWS

problem. We can verifythis by noting that the first prolongation of v is, as the reader is urged to verify,

Therefore,

pr(l)vCL) = 0+ 0+ CC1-

uin!c- 2ux)

R 1-

On theother hand, since X

LDxCX) =

= u and U

ui

= -ux)l- u1.

= x,

)1- u1 DxCu) = )1- u1ux,

so that EqoationC30.20) is satisfied.

III

In the last chapter, we studied the symmetries of the DEs in some detail. This chapter introduces us to a particular DE that arises from a variational problem, namely, the Euler-Lagrange equation. The natural question to ask now is, How does the variational symmetry manifest itself in the Euler-Lagrange equation? Barring some technical difficulties, we note that for any change of variables, if u = f(x) is an extremal of the variational problem L[u], then il = lei) is an extremal of the variational problem [[ il]. In particular, if the change is achieved by the action of the variational symmetry group, (i, il) = g . (x, u) for some g E G, theu [[il] = L[il], and g . f is also an extremal of L. We thus have

30.2.5. Theorem. If G is the variational symmetry group ofa functional, then G is also the symmetry group ofthe associated Euler-Lagrange equations. Symmetries ofthe Euler-Lagrange equations are not necessarily the symmetries ofthe corresponding variational probleml

The converse is not troe! There are symmetry groups of the Euler-Lagrange equations that are not the symmetry group of the variational problem. Problem 30.8 illustrates this for p = 3, q = I, and the functional

L[u] =

~fff(ul-u; -u~)dxdydt,

(30.21)

whose Euler-Lagrange equation is the wave equation. The reader is asked to show that while the rotations and Lorentz boosts ofTable 29.3 are variational symmetries, the dilatations and inversions (special conformal transformations) are not. We now treat the case of p = 1 = q, whose Euler-Lagrange equation is an ODE. Recall that the koowledge of a symmetry group of an ODE led to a reduction in the order of that ODE. Let us see what happens in the present case. Suppose v = xax + ua" is the infinitesimal generator of a l-parameter group of variational symmetries of L. By an appropriate coordinate transformation from (x, u) to (y, w), as in Section 29.5, v will reduce to a/aw, whose prolongation is alsoa/aw. In terms of the new coordinates, Equation (30.20) will reduce to

991

30.2 SYMMETRY GROUPS OF VARIATIONAL PROBLEMS

a

L/ aw = 0; i.e., the new Lagrangian is independent of w, and the Euler-Lagrange equation (30.14) becomes

(30.22) Therefore, the expression in the brackets is some constant A (because D y is a total derivative). Furthermore, if we introduce v = w y , the expression in the brackets becomes the Euler-Lagrange equation of the variational problem t[v] =

f

t(y, v(n-I» dy,

where

,

L(y,

v

(n-I)

)

-

= L(y, wy , ••. , wn ) ,

and every solution w = f (y) of the original (2n )th-order Euler-Lagrange equation corresponds to the (2n - 2)nd-order equation

"

at n-I . at + L(-Dy)J- = ay j=1 aVj

JE(L) = -

A.

(30.23) .

Moreover, this equation can be written as the Euler-Lagrange equation for tA[v] =

Lagrange multiplier

f

[t(y, v(n-I» - AV] dy,

and A can be thought of as a Lagrange multiplier, so that in analogy with the multivariable extremal problem," the extremization of tA[v] becomes equivalent to that of t[v] subject to the constraint J v dy = O. We summarize the foregoing

discussion in thefollowingtheorem. 30.2.6. Theorem. Let p = I = q, and L[u] an nth-order variational problem with a i-parameter group of variational symmetries G. Then there exists a oneparameter family of variational problems t A [v] of order n - I such that every solution of the Euler-Lagrange equationfor L[u] can befound by integrating the solutions to the Euler-Lagrange equation for tA[v]. Thus, we have the following important result:

30.2.7. Box. A i-parameter variational symmetry of a functional reduces the order ofthe corresponding Euler-Lagrange equation by two.

This conclusion is to be contrasted with the symmetry of ODEs, where each 1parameter group of symmetries reduces the order of the ODE by 1. It follows from 4See [Math 70. pp- 331-341] fora discussionof Lagrange multipliers andtheiruse in variational techniques, especiallythose used in approximating -solutions of the Schrodinger equation.

992

30. CALCULUS OF VARIATIONS, SYMMETRIES, AND CONSERVATION LAWS

Box 30.2.7 that the ODEs of order 2n derived from a variational problem-the Euler-Lagrange equation-are special.

30.2.8. Example. A first-order variational problem witb a I-parameter group ofsymmetries can be integrated out. By transfonning to a new coordinate system, we can always asswne that the Lagrangian is independent of the dependent variable (see Proposition 29.5.1). TheEuler-Lagrange equation in tbis casebecomes

st. au '-or-'

az,

O=lE(L) = - - Dx -

aux

=}

aL

-(x, ux) = A. aux

=0

Solving this implicit relation, we get Ux function ofx (and A).

= F (x, ).,); which can be integrated to give u as a III

The procedure can be generalized to r-parameter symmetry groups, but the order carmotbe expected to be reduced by 2 unless the group is abelian. We shall not pursue this matter here, but ask the reader to refer to Problem 30.9.

30.3 Conservation Laws and Noether's Theorem A conservedphysicalquantityis generallydefinedas a quantitywhosefluxthrough

any arbitrary closed surface is equal to (the negative of) the rate of depletion of the quantity in the volume enclosed. This statement, through the use of the divergencetheorem,translatesinto a relation connectingthe time rate of changeof the densityand the divergenceof the current ~orresponding to the physicalquantity. Treating time and space coordinates as independent variables and extending to p independentvariables, we have the following: current density and conservation law

30.3.1. Definition. A conservation law for a system of differential equations
= ovalidfor all solutions u =

f(x)

is called current density.

constant ofthe motion, orfirst integral of a system of ODEs

For p = I = q, i.e., for a system of ODEs, a conservation law takes the form DxJ(x, u(n») = 0 for all solutions u = f(x) of the system. This requires J (x, u(n») to be a constant,i.e., that J (x, u(n») be a constant or the motion, or, as it is sometimes called, the first integral of the system. In order to understand conservation laws, we need to get a handle on those conservationlaws that are trivially satisfied.

trivial conservation iaw of the first kind

30.3.2. Definition. If the current density J itself vanishes for all solutions u = fix) of the system
30.3 CONSERVATION LAWS AND NOETHER'S THEOREM

993

To eliminate this kind of triviality. one solves the system and its prolongations

t. (k)(x. u(n» = 0 for some of the variables u~ in terms of the remaining vatiables and snbstitutes the latter whenever they occur. For example. one can differentiate the evolution equation u, = Ftx, u(n»_in which u(n) have derivatives with respect to x only-with respect to t and x sufficient number of times (this is what is meant by "prolongation" of the system ofequations) and solve for all derivatives of u involving time. Then. in the conservation law. substitute for any such derivatives to obtain a conservation law involving only x derivatives of u. 30.3.3. Example. The current density J1 = conservedfor the systemof first-orderDEs Ut

= ve,

Ux

=

(!uF + !u~. -u,ux) is easily seen to be

Vt.

(!ui !v;. -uxvx), which

By eliminating all the time derivatives in Jj , we obtain J2 = + is also conserved. However, the difference between these two currents,

J

= Jr - J2 = (!u; - !v;, UxVx - UtUx),

satisfiesa trivial conservation law of the first kind, because the componentsof J vanish on the solutions of the system. II trivial conservation law ofthe second kind; null divergence

30.3.4. Definition. If the current density J satisfies D . J = 0 for all functions u = f(x), even if they are not solutions of the system of DEs, the divergence identity is called a trivial conservation Taw of the second kind. In this case J is called a null divergence.

If we treat J; as the components of a (p - I)-form w. so that the exterior derivative dw is the divergence of J (times a volume element). then the triviality of the conservation law for J is equivalent to the fact that w is closed. By the converse of the Poincare lemma. there must be a (p - 2)-form 1} such that w = dn. In the context of this chapter. we have the following theorem. 30.3.5. Theorem. Suppose J = (1J(x. u(n» •...• Jp(x. u(n») is a p-tuple of functions on X x Then J is a null divergence if and only if there exist smooth functions Akj(X. u(n», i. k = I •...• p, antisymmetric in their indices, such that

u».

p

h

= LDjAkj.

j = I .... • p.

(30.24)

)=1

trivial conservation law and the equivalence oftwo conservation laws

30.3.6. Definition. We say that D . J = 0 is a trivial conservation law if there exist antisymmetric smooth functions Akj (x, u(n» satisfying Equation (30.24) for all solutions of the system of DEs t.(x. u(n» = O. Two conservation laws are equivalent if they differ by a trivial conservation law. We shall not distinguish between conservation laws that are equivalent. It turns out that to within this equivalence. some systems of DEs have current densities J

994

30. CALCULUS OF VARIATIONS, SYMMETRIES, AND CONSERVATION LAWS

such that for some I-tuple Q = (Qt, ... , Q/),

(30.25)

where {Qv} are smooth functions of x, u, and all derivatives of u. characteristic form of a conservation law

30.3.7. Definition. Equation (30.25) is called the characteristic fiarm ofthe conservation law for the current density J, and the l-tuple Q, the characteristic ofthe conservation law.

Wearenow in a positionto prove the celebrated Noether's theorem. However, we first need a lemma.

L1=1 Xiajax i + L~~l U"ajau" where Xi

30.3.8. Lemma. Let v = are functions of x and u. Let

and U"

p

Q"(x, u(l)) sa U"(x, u) - LXi(x, u)ur, .

a

= 1, ... .q,

i=1

Then p

pr(n)v = pr(n)vQ + LXi o..

(30.26)

i=1

where

The sum over J extends over all multi-indices with 0 term being simply vQ.

s IJ I s n, with the IJ I = 0

Proof Substitute Q" in the definition of U'J as given in Theorem 29.3.5 to obtain p

U'J = DJQ"

+ LXiuj,/, ;=1

where snm)

ug = Q" + L1~1 Xi ur = U". It follows that (with J = 0 included in the

=Dj by Proposition 29.3.4

and the lemma is proved.

D

30.3 CONSERVATION LAWS ANO NOETHER'S THEOREM

The celebrated Noether's theorem connecting symmetries to conservation laws

995

30.3.9. Theorem. (Noether's theorem) Let p

q

v= LX'8/8x'

+ L uu8/8uu a=l

i=1

be the infinitesimalgenerator ofa local s-parameter group ofsymmetries G ofthe variational problem L[u] = J L(x, u(n» dPx. Let P

QU(x, u(l» '" UU(x, u) - L

X'(x, u)uf,

i=1

Then there exists a p-tuple J = (11, ... , Jp ) such that q



J = L QUlEu(L)

(30.27)

a=l

is a conservation law in characteristic form for the Euler-Lagrange equation lEu(L) = O. Proof We use Lemma 30.3.8 in the infinitesimal criterion of the variational symmetry (30.20) to obtain

0= pr(n)v(L) + LD . X p

p

= pr(n)vQ(L) + LXiD,L+LLDiXi ;=1 p

= pr(n)vQ(L) + L

(30.28)

i=1

Di(LX') = pr(n)vQ(L) + D· (LX).

i=1

Using the definition of pr(n)vQ and the identity

(DjS)T = Dj(ST) - SDjT, we can commute DJ = DiI ... Dik past QU one factor at a time, each time introducing a divergence. Therefore,

'" U----a=L.JQ 8L ' " U(-D)J----a+ 8L D. A pr(n) vQ(L)=L.JDJQ a,J au] a,J au] q

= L

QUlEu(L) + D· A,

01=1

where A = (AI, ... , A p ) is some p-tnple of functions depending on L, the QU's, and their derivatives, whose precise form is not needed here. Combining this with Equation (30.28), we obtain q

0= L

QUlEu(L) + D· (A + LX).

a=l

Selecting J = -(A + LX) proves the theorem.

D

996

30. CALCULUS OF VARIATIONS, SYMMETRIES, AND CONSERVATION LAWS

Amalie Emmy Noether (1882-1935), generally considered the greatest of all female mathematicians up to her time, was the eldest child of Max Noether, research mathematician and professor at the University of Erlangen, and Ida Amalia Kaufmann. Two of Em my's three brothers were also scientists. Alfred, her junior by a year, earned a doctorate in chemistry at Erlangen. Fritz, two and a half years younger, became a distinguished physicist; and his son, Gottfried, became a mathematician. At first Emmy Noether had planned to be a teacber of English and French. From 1900 to 1902 she studied mathematics

and foreignlanguagesat Erlangen.Then in 1903she startedher specializationin mathematics at the University of Gottingen. At both universities she was a nonmatriculated auditor at lectures, since at the tum of the century women could not be admitted as regular students. In 1904 she was permitted to matriculate at the University of Erlangen, which granted her the Ph.D., summa cum laude, in 1907. Her sponsor, the algebraist Gordan, strongly influenced her doctoral dissertation on algebraic invariants. Her divergence from Gordan's viewpoint and her progress in the direction of the "new" algebra first began when she was exposed to the ideas of Ernst Fischer, who came to Erlangen in 1911. In 1915 Hilbert invited Emmy Noether to Gottingen. There she lectured at courses that were given under his name and applied her profound invariant-theoretic knowledge to the resolution of problems that he and Felix Klein were considering. Inspired by Hilbert and Klein's investigation into Einstein's general theory of relativity, Noether wrote her remarkable 1918 paper in which both the concept of variational symmetry and its connection with conservation laws were set down in complete generality. Hilbert repeatedly tried to obtain her an appointment as Privatdozent, but the strong prejudice against women prevented her habilitation until 1919. In 1922 she was named a nichtbeamteter ausserordentlicher Professor ("unofficial associate professor"), a purely honorary position. Subsequently, a modest salary was provided through a Lehrauftrag (vreaching appointment") in algebra. Thns she taught at Giittingen (1922-1933), interrupted only by visiting professorships at Moscow (1928-1929) and at Frankfurt (summer of 1930). In April 1933 she and other Jewish professors at Gottingen were summarily dismissed. In 1934 Nazi political pressures caused her brother Fritz to resign from his position at Breslau and to take up duties at the research institute in Tomsk, Siberia. Through the efforts of Hermann Weyl, Emmy Noetherwas offered a visiting professorship at Bryn Mawr College; she departed for the United States in October 1933. Thereafter she lectured and did research at Bryn Mawr and at the Institute for Advanced Studies, Princeton, but those activities were cut short by her sudden death from complications following surgery. Emmy Noether's most important contributions to mathematics were in the area of abstract algebra. One of the traditional postulates of algebra, namely the commutative law of multiplication, was relinquished in the earliest example of a generalized algebraic structure, e.g., in Hamilton's quatemion algebra and also in many of the 1844 Grassmann algebras. From 1927 to 1929 Emmy Noether contributed notably to the theory of representations, the object of which is to provide realizations of noncommutative algebras by means of matrices, or linear transformations. From 1932 to 1934 she was able to probe profoundly

30.4 APPLICATION TO CLASSICAL FIELO THEORY

997

intothestructure of noncommutative algebras by means of her concept of the verschriinktes ("cross") product.

Emmy Noether wrote someforty-five research papers andwas an inspiration to many future mathematicians. The so-called Noether school included such algebraists as Hasse and W. Schmeidler, with whom she exchanged ideas and whom she converted to her own special point of view.Shewasparticularly influential in the work of B. L. van der Waerden, who continued to promote herideas after herdeath andto indicate the many concepts for whichhe wasindebted to her.

30.4 Application to Classical Field Theory It is clear from the proof of Noether's theorem that if we are interested in the conserved current, we need to find A. In general, the expression for A is very complicated. However, if the variational problem is of first order (which in most . cases of physical interest it is), then we can easily find the explicit form of A, and, consequently the conserved current J. We leave it for the reader to prove the following: 30.4.1. Corollary. Let v = Lr~l xia/ax i + L~~l uaa/au a be the infinitesimal generator of a local I-parameter group of symmetries G of the first-order variational problem L[u] = J L(x, u(l) dt'x, Then 5

i = 1, .. o,p form the components of a conserved current for the Euler-Lagrange equation IEa(L) = O. This corollary can be applied to most DEs in physics detivable from a Lagrangian. We are interested in partial DEs stodied in classical field theories. The case of ODEs, stodied in point mechanics, is relegated to Problem (30.11). Fitst consider spacetime translation vi = ~ij j, where we have introduced the Lorentz mettle ~ij to include non-Euclidean cases. 10 order for v' to be an infinitesimal vatiational symmetry, it has to satisfy Equation (30.20), which in the case at hand, reduces to vi (L) = 0, or aiL = O.

a

30.4.2. Box. In orderfor a variational problem to be invariant under spacetime translations, its Lagrangian must not depend explicitly on the coordi-

nates. 5Wehavemultiplied Jj by a negativesign to conform to physicists'convention.

998

30, CALCULUS OF VARIATIONS, SYMMETRIES, AND CONSERVATION LAWS

If spacetime translation happens to be a symmetry, then Xi -+ ryi}, and the (double-indexed) conserved current, derived from Corollary 30.4.1, takes the form

,. -f. au a aL

..

T'J = L. - - - - - ry'J L u=l

energy momentum current density

axj auf

.

Using Greek indices to describe space-time coordinates, and Latin indices to label the components of lltq , we write TfLV =

q'

L

q'

L

at/>J aL _ ryfLV L sa ryfL rr at/>~ aL _ ryfL VL, '-1 ax" '-I _ . ax at/>Jv Jr a",J o/v J

(30.29)

where we changed the dependent variable u to t/> to adhere to the notation used in the physics literature. Recall that t/>t es at/>j lax v . TfLV is called the energy, momentum current density. The quantity T fLV , having a vanishing divergence, is really a density, just as the continuity equation (vanishing of the divergence) for the electric charge involves the electric charge and current densities. In the electric case, we find the charge' by integrating the charge density, the zeroth component of the electric 4-current density. Similarly, we find the "charge" associated with TfLV by integrating its zeroth component. This yields the energy momentum 4 vector:

pv =

Iv TO vd3x .

We note that

, where we have used the three-dimensional divergence theorem. By taking S to be infinite, and assuming that T i v -+ 0 at infinity (faster than the element of area da, diverges), we obtain d P" [dt = 0, the conservation of the 4-momentum.

30.4.3. Example. A relativistic scalar field of mass m is a l-component field satisfying theKlein-Gordan equation, whichis, asthereader maycheck, theEuler-Lagrange equation of

Theenergymomentum current for the scalar fieldis foundto be

30.4 APPLICATION TO CLASSICAL FIELD THEORY

999

Notethat Tf.L V is symmetric under interchange of its indices. Thisis a desired feature of the energy momentum current that holds for the scalar fieldbutis not satisfied in general, as Equation(30.29) indicates. Thereaderis urgedto showdirectlythat a"T"v = 0 = avT"v, i.e., thatenergy momentumis conserved. III

To go beyond translation, we consider classical (nonquantized) fields 6{t/Jj} j:::l' which, as is the case in most physical situations, transform among themselves as the rows of the ath irreducible representation of a Lie group G that acts on the independent variables. Under these circumstances, the generators of the symmetry are given by Equation (27.59): k

(a)

a + 8ijX v (x; I;)-a av'

:Dij(l;) = 'rij (I;)t/J (X)-k at/J

(30.30)

x

where v labels the independent variables. Corollary 30.4.1 now gives the conserved current as

J!j

= {X" (x; I;)t/Je(x) aLk - X"(x; I;)L} 8ij - t/Jk(x) a~ ~j)(I;), at/Jv at/J~

where summation over repeated indices is understood with I I :s: v :s: p. We can rewrite this equation in the form

J" = {X"(X; I;)t/Jk(x) aL _ X"(x; I;)L} 1 _ t/Jk(x) aL v

at/J~

at/Je

:s:

k

:s:

n a and

~(a)(I;), (30.31)

where J" and ~(a)(I;) are n« x n a matrices whose elements are J!j and 'rij) (1;), respectively, and 1 is the unit matrix of the same dimension. We note that the conserved current has a coordinate part (the term that includes X" and multiplies the unit matrix), and an "intrinsic" part (the term with no X") represented by the term involving ~(a) (1;). If the field has only one component (a scalar field), then ~(a) (1;) = 0, and only the coordinate part contributes to the current.

The current J" acqnires an extra index when a component of I; is chosen. As a concrete example, consider the case where G is the rotation group in RP. Then a typical component of I; will be ~ po , corresponding to a rotation in the pa -plane, and the current will be written as J";pu. These extra indices are also reflected in X", as that too is a function of 1;:

a

X" (x; ~PU) ax" orbital angular momentum and intrinsic spin

= x Pau -

XU ap

=}

X" (x; ~pU)

= x" 8"u

- XU 8"p.

The volume integral of JO;pu will give the components of angular momentum. When integrated, the term multiplying 1 becomes the orbital angular momentum, and the remaining term gives the intrinsic spin. The conservation of J";pu is the

6The reader notes thatthe superscript a, which labeledcomponents of the independent variable u, is now the label of the irreducible representation., The components of the dependent variable (nowdenoted by ¢) arelabeledby i-

1000

30. CALCULUS OF VARIATIONS, SYMMETRIES, ANO CONSERVATION LAWS

statement of the conservation of total angular momentum. The label O! denotes various representations of the rotation group. If p = 3, then O! is simply the value of the spin. For example, the spin-~ representation corresponds to O! = ~,and or with a labeling the three different "directions" of rotation? If the field is a scalar, :rCa)(e) = 0, and the field has only an orbital angnlar momentum.

30.5 Problems 30.1. Show that the derivative of a linear map from one Hilbert space to another is the map itself. 30.2. Show that a complex function f : iC ::J g --+ iC considered as a map f:]R2::J g --+ ]R2 is differentiable iffit satisfies the Cauchy-Riemann conditions. Hint: Consider the Jacobian matrix of f, and note that a linear complex map T : iC --+ iC is necessarily of the form T(z) = AZ for some constant A E C. 30.3. Show that

8Ey,i [U] ==:::'(x) = -ai8(X - y).

8u

30.4. Show that the first functional derivative of L[u] '" J~2 using Equation (30.9), is JE (L).

vi 1 + ui dx, obtained

30.5. Show that for the proper time of special relativity 8L[x] x" 8x(s) = (1 - x})3/2'

Use this to show that the contribution of the second variational derivative to the Taylor expansion of the functional is always negative. 30.6. Show that the first prolongation of the Lorentz generator v = uax

+ xau is

30.7. Verify that rotation in the xu-plane is a symmetry of the arc-length variational problem (see Example 30.1.13). 30.8. Show that V4, V6, and V7 ofTable 29.3 are variational symmetries of Equation (30.21), but vs, vg, V9, and VlO are not. Find the constant c (if it exists) such that Vs + cuau is a variational symmetry. Show that no linear combination of inversions produces a symmetry.

30.5 PROBLEMS

Kepler problem

1001

30.9. The two-dimensional Kepler problem (for a unit point mass) starts with the functional

(a) Show that L is invariant under t translation and rotation in the xy-plane. (b) Find the generators of t translation and rotation in polar coordinates and conclude that r is the best choice for the independent variable. (c) Rewrite L in polar coordinates and show that it is independent of t and (d) Write the Euler-Lagrange equations and integrate them to get as an integral over r.

e

e.

30.10. Prove Corollary 30.4.1. 30.11. Consider a system of N particles whose total kinetic energy K and potential energy U are given by N

! L:mal:i"12 ,

K(x) =

a=l

where x" = (x a, is of the form L[x] =

i:

U(t,x) = L:kapl~ _xPI- 1 , a#

s". za) is the position of the ath particle. The variational problem

ui. x, x) dt =

i:

[K(x) - U(t, x)] dt.

(a) Show that the Euler-Lagrange equations are identical to Newton's second law of motion. (b) Write the infinitesimal criterion for the vector field

a v = ,(t, x)-

"[a (t, x)-a a +

at + L.., a

~

ax

~ a (t,

a x)---"

a a] ay + ~ (t, x)----a az

to be the generator of a I-parameter group of variational symmetries of L. (c) Show that the conserved "current" derived from Corollary 30.4.1 is N

T = L:ma (~aia

+ ~aya + ~aia) -

,E,

Cl=l

where E = K + U is the total energy of the system. (d) Find the conditions on U such that (i) time translation, (ii) space translations, and (iii) rotations become symmetries of L. In each case, compute the corresponding conserved quantity. 7Only in three dimensions can one label rotations with a single index. This is because each coordinate plane has a unique direction (by the use of the right-hand rule) perpendicular to it that can be identified as the direction of rotation.

1002

30. CALCULUS OF VARIATIONS, SYMMETRIES, ANO CONSERVATION LAWS

30.12. Show that the Euler-Lagrange equatiun of

is the Klein-Gordan equation. Verify that TILv = 8 IL
conserved.

Additional Reading The first three books cited below discuss what we have not covered in our study of Lie groups and DEs, namely "generalized symmetries," whereby the generators of synuuetries are allowed to be not only functions of the independent and dependent variables, but also of the derivatives of the dependent variables. This leads to a more general version of Noether's theorem than presented in this chapter.

1. Blurnan, G. and Kurnei, S. Symmetries and Differential Equations; SpringerVerlag, 1989. 2. Olver, P. Application of Lie Groups to Differential Equations, SpringerVerlag, 1986. 3. Stephani, H. Differential Equations: Their solutions using symmetries, Cambridge University Press, 1989. 4. All books on relativistic quantum field theories have a discussion of symmetries and conservation laws. See, for example, Weinberg, S. The Quantum Theory ofFields (2 volumes), Cambridge University Press, 1995.

Bibliography

[Abra 85] Abrabam, R. and Marsden, J. Foundations Wesley, 1985.

0/ Mechanics,

2nd ed., Addison-

[Abra 88] Abrabam, R., Marsden, J., and Ratiu, T. Manifolds, Tensor Analysis, andApplications, 2nd ed., Springer-Verlag, 1988. [Axle 96] Axler, S. Linear Algebra Done Right, Springer-Verlag, 1996. [Baru 86] Barut, A. and Raczka, R. Theory World Scientific, 1986.

0/ Group Representations and Applications,

[Birk 77] Birkboff, G. and MacLane, S. Modern Algebra, 4th ed., Macntillan, 1977. [Birk 78] Birkboff, G. and Rota, G.-C. Ordinary Differential Equations, 3rd ed., Wiley, 1978. [Bish 80] Bishop, R. and Goldberg, S. Tensor Analysis on Manifolds, Dover, 1980. [Blum 89] Bluman, G. and Kumei, S. Symmetries and Differential Equations, SpringerVerlag, 1989.

[Bocc 90] Boccara, N. Functional Analysis, Academic Press, 1990. [Boer 63] Boerner, H. Representation a/Groups, North-Holland, 1963. [Batt 82] Batt, R. and To, L. Differential Forms in Algebraic Topology, Springer-Verlag, 1982. [Chev 46] Chevalley, C. Theory a/Lie Groups, Princeton University Press, 1946. [Choq 82] Choquet-Brohat, Y.,DeWitt-Marette, C., and Dillard-Bleick, M. Analysis, Manifolds, and Physics, 2nd ed., North-Holland, 1982. [Chur 74] Churchill, R. and Verhey, R. Complex Variables and Applications, 3rd ed., McGraw-Hill, 1974. [Cour 62] Courant, R. and Hilbert, D. Methods science, 1962.

0/ Mathematical Physics,

vol. I, Inter-

1004

BIBLIOGRAPHY

[Denn 67] Dennery, P., and Krzywicki, A. Mathematics for Physicists, Harper and Row, 1967.

[DeVi 90] DeVito,C. Functional Analysis and Linear Operator Theory, Addison-Wesley, 1990. [DeVr 94] DeVries, P.A First Course in Computational Physics, Wiley, 1994. [Eeon 83] Economou, E. Green's Functions in Quantum Physics, SpringerVerlag, 1983. [Elli 79] Elliott, J., and Dawber, P. Symmetry in Physics (2 volumes), Oxford University Press, 1979.

[Plan 89] Flanders, H. Differential Forms with Applications to Physical Sciences, Dover, 1989. [Fo1l95] Folland, G. Introduction to Partial Differential Equations, 2nd ed., P.rluceton

University Press,1995. [Frie 82] Friedman, A. Foundations ofModem Analysis, Dover, 1982. [Fu1191] Fulton, W. and HarrIs, J. Representation Theory, Springer-Verlag, 1991. [Gill72] Gillispie, C., ed. Dictionary of Scientific Biography, Charles Scribner's, New York, 1970.

[Gilm 74] Gilmore, R. Lie Groups, Lie Algebras, and Some of Their Applications, Wiley, 1974. [Glim 87] Glimm, J. and Jaffe, A. Quantum Physics, 2nd ed., Springer-Verlag, 1987. [Grad 65] Gradshleyn, I. and Ryzhik, I. Table ofIntegrals, Series, and Products, Academic Press, 1965. [Greu 75] Greub, W. Linear Algebra, 4th ed., Springer-Verlag, 1975. [Halm 58] HaImos, P. Finite Dimensional Vector Spaces, 2nd ed., Van Nosttand, 1958. [Halm 74] HaImos, P. Naive Set Theory, Springer-Verlag, 1974. [Hame 89] Harnermesh, M. Group Theory and Its Application to PhysicalProblems, Dover, 1989. [Hass 00] Hassani, S. Mathematical Methods for Students ofPhysics and Related Fields, Springer-Verlag, 2000. [He1l67] Hellwig, G. Differential Operators of Mathemotical Physics, Addison-Wesley, 1967. [Hild 87] Hildebrand, F. Introduction to Numerical Analysis, 2nd ed., Dover, 1987. [Hi1l87] Hildebrand, F. Statistical Mechanics, Dover, 1987. [Jack 75] Jackson, 1. Classical Electrodynamics, 2nd ed., Wliey, 1975. [Jorg 82] Jorgen, K. Linear Integral Operators, Pittnan, 1982. [Ke1l85] Kelley, J. General Topology, Springer-Verlag, 1985. [Koba 63] Kobayashi, S. and Nomizn, K. Foundations ofDifferential Geometry, vol. I, Wiley, 1963. [Lang 85] Lang, S. Complex Analysis, 2nd ed., Springer-Verlag, 1985.

BIBLIOGRAPHY

1005

[Lorr 88] Lorrain, P., Corson D., andLorrain, F. Electromagnetic Fields and Waves, 3rd ed., W. H. Freeman, 1988.

[Mack 68] Mackey, G. Induced Representations, Benjamin, 1968. [Marl 80] Marion, J. and Heald, M. Classical Electromagnetic Radiation, 2nd ed., Academic Press, 1980. [Malh 70] Mathews, J. and Walker, R. Mathematical Methods of Physics, 2nd ed., Benjamin, 1970. [Mess 66] Messiah, A. Quantum Mechanics (2 vnlnmes), Wiley, 1966. [Mill68] Miller, W. Lie Theory and Special Functions, Academic Press, 1968.

[Misn 73] Misner, C., Thome,K., andWheeler, J. Gravitation, Freeman, 1973. [Mars 53] Morse, P. and Feshbach, M. Methods of Theoretical Physics, McGraw-Hill, 1953. [Naka 90] Nakahara, M. Geometry, Topology, and Physics, Adam Hilger, 1990.

[DIve86] Olver, P. Application ofLie Groups to Differential Equations, Springer-Verlag, 1986. [Reed 80] Reed, M. and Simon, B. Functional Analysis (4 vnlnmes), Academic Press, 1980. [Rich 78] Richtmyer, R. Principles ofAdvanced Mathematical Physics, Springer-Verlag, 1978. [Roac 70] Roach, G. Green's Functions, Van Nostrand, 1970. [Rotm 84] Rotman, J. An Introduction to the Theory ofGroups, 3rd ed., Allyn and Bacon, 1984. [Rndl 74] Rndln, W. Functional Analysis, McGraw-Hill, 1991. [Sann 89] Sannders, D. The Geometry ofJet Bundles, Cambridge University Press,

1~89.

[SImm 83] Simmons, G. Introduction to Topology and Modem Analysis, Krieger, 1983. [SImm 92] Simmons, G. Calculus Gems, McGraw-Hill, 1992. [Stak 79] Stakgold, I. Green's Functions and Boundary Value Problems, Wiley, 1979. [Step 89] Stephani, H. Differential Equations: Their solutions using symmetries, Cambridge University Press, 1989.

[Tric 55] Tricomi, F. Vorlesungen uber Orthogonalreihen, Springer, 1955.

[Vara 84] Varadarajan, V. Lie Groups, Lie Algebras and Their Representations, SpringerVerlag, 1984. [Wald 84] Wald, R. General Relativity, University of Chicago Press, 1984. [Warn 83] Wamer, F. Foundations ofDifferentiable Manifolds and Lie Groups, SpringerVerlag, 1983. [Wats 52] Watson, G. A Treatise on the Theory of Bessel Functions, 2nd ed., Cambridge University Press, 1952. [Wein 95] Weinberg, S. The Quantum Theory ofFields (2 volnmes), Cambridge University Press, 1995. [Wign 59] Wigner, E. Group Theory, Academic Press, 1959.

[Zeid 95] Zeidler, E. Applied Functional Analysis, Springer-Verlag, 1995.

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

.

Index

J\bel,63, 177, 183,233,416,462,653 biography, 662 abelian group, 654 abelian Lie algebra, 834

absolute convergence, 252 additive identity, 20 adjoint formal, 560 adjoint action, 827 adjoint algebra, 840 adjoint Be, 563 adjoint DO, 364-367 adjoint Green's function, 566 adjoint map, 825 adjoint of a matrix, 87

adjoint of anoperator, 61 adjoint representation, 679 character, 684 affine group, 876 algebra, 41-45 definition, 41

derivation of an,42 dimensionof an, 41 homomorphism, 43 ideal,44

synunettic, 737 tensor, 731 algebraic equation, 937 synunetry gronp of, 937 analytic cootinuation, 302-309 analytic function, 228-235 definition, 233 derivative, 228

derivatives as integrals. 249 entire, 233 poles of, 273 roots (zeros) of, 262 angular momentum, 831, 869 addition theorem, 866 eigenvalues, 337 operator, 331

construction of eigenvalues, 334 annihi1ator, 39 anti-hermitian operator, 63 anticommutator, 59

antisymmetric representation, 679 antisymmetrizer, 739 arc length, 890

associated Legendre functions, 341 associative algebra, 41

isomorphism, 43 of operators, 49--76 quaternions, 44

averaging operator, 71

structure constants, 43

azimuthal symmetry, 343

atlas, 765 automorphism, 36

1008

INDEX

backward difference operator. 70, 384-387 Baker-Campbell-Hausdorff, 59 Banach space, 148 basis, 22 dual of a, 38, 729 oriented,745 orthouormal, 26 transformation matrix, 91 Becquerel, 799 Bernoulli, 423, 981 Bessel, 177, 620, 982 biography, 422

Bessel equation Liouvillesubstitution, 514

periodic, SIS separated, 510 unrnixed,564 boundery functionals, 559 boundery point, 459 boundery value problem, 560, 583 DiricWet,591 Neumann, 591

numerical solution, 383 hounded operator, 453-457 continnity,454 branch cut, 296 Brouwer, 850

bundle of tensors, 786 BWHB theorem, 853

Bessel functiun, 423-426

asymptotic behavior, 443 large argument, 444 large order, 443 confluent hypergeomettic, 424 generating function, 440 integral representation of, 438 modified, 425 second kind, 425 oscillation of, 363 second kind, 424 third kind, 425

Bessel functions spherical, 542 Bessel imaginary (bei), 539 Bessel inequality, 149 Bessel real (ber), 539 beta function, 309-312 definition, 310 Bianchi identity, 889 bijective map, 6

binary operation, 7 binomial theorem, 13 Birkhoff's theorem, 925 Bolzano,l1

Balzano-Weierstrass theorem, 461 Boole,704 boundary conditions Dirichlet, 591 homogeneous, 559 mixed, 564 Neumann, 591

canonical basis, 747, 802 second order tinear DE, 404 canonical coordinates, 802 canonical transformation, 802, 806 Cantor, 463, 739, 799 biography, 10 Cantorse!, 12 cardina1ity, 12 Cartan,798,841,942 biugraphy, 743 Cartan mettic tensor, 843, 862 Cartan subalgebra, 845 Cartan's lemma, 743

Cartesian product, 2 Casimir operator, 861-863, 868-870, 872, 873 Cauchy, 96, 271, 296, 416, 478, 528, 588, 653,891 biography, 233 Cauchy data, 584 Cauchy integral formula, 245 for operators, 482 Cauchy problem, 584 ill-posed, 590 Cauchy sequence, 10, 146 Cauchy-Goursat theorem, 242 Cauchy-Riemann conditions, 230 differentiability, 232 Cayley, 704, 744 center of a group, 658

central difference operator, 70

INDEX

centralizer, 658 Champollion, 198

compact Lie group representation, 845-855

character adjoint representation, 684

compact operator, 464-466 definition, 464 spectturn,467-473

compound, 683, 685 conjugacy class, 684 group and its subgroup, 692 of a representation, 683 simple, 683, 685 character table, 691

character table for S2, 694 character table for S3, 694 characteristic hypersorface, 585-589 definition, 586 characteristic polynomial, 115 linear DEs, 376 chart, 764 Chebyshev, 588 Chebyshev polynomials, 185 Chevalley's theorem, 862 Chevalley, biography, 864 Christoffel symbol, 890 Christoffel, biography, 890 circnit matrix of a DE, 403 circular heat-conducting plate, 537 Clebsch, 178, 704 biography, 702 Clebsch-Gordan coefficients, 705, 706 decompo~tion,701-705,721,866

series, 703, 706 Clifford, 744 closed form, 797 closed snbset, 459 closure, 459

codomain,S cofactor, 95 commntative algebra, 41

commutative group, 654 commutative Lie algebra, 834 commutator diagona1izability, 124 commutator subgroup, 658 commutators, 55 compact Lie algebra, 843

compact resolvent, 508 compact set, 458-463 compact subset, 461

compact support, 165n,801

comparison theorem, 361-363 complement of a set, 2 complete 0.0. sequence, 149 cornpletenessrelatipn,68,150 complexcoordinate space, 21 complex exponential function, 234 complex FOillEs, 401-402 complex function, 227 integration, 241 complex GL(V), 822 complexplane contour in the, 242 curvein the, 242 multiply connected region, 245 path in the, 242 simply connected region, 245 complex potential, 236

complex series, 252 complex SOLDE, 404-410

complexvector space, 20 composition of maps, 6 compound character, 685 conducting cylindrical can, 535 confluenthypergeometric function definition, 420 integral representation of, 437 conformal gronp, 912 in 2 dimensions, 912 conformal Killing vector, 911 conformal map, 236-241 definition, 237 translation, 238

conformal transformation, 911 special, 912

conic sections, 132 conjngacy class, 661, 684 conjngate,661 conjugate subgronp, 657

1009

1010

INDEX

conjugation of operators, 61-63 conjunct, 560, 563,576 connection coefficients, 886, 900

cyclic permutation, 666 cyclic subgronp, 657

conservation law,855, 992

d' Alembert, 984 biography, 330 Darboux, 744, 942 Darbonx inequality, 245 Darbonx theorem, 802 de Broglie, 807 Dedekind, 11, 711,739,891 degeneracy, 697 energy, 605 degenerate kernels, 501-504 delta function, 160, 592 and step function, 163

characteristic, 994 trivial, 993 trivial of the first kind, 992 trivial of the second kind, 993

conserved current density, 992 constant of themotion, 992 constrained systems,8040

continuous index, 159 contraction, 735 contravariant degree, 731

contravariant tensor, 731 convexsubset, 474 convolution theorem, 223 coordinate curve, 774 coordinate frame, 774 coordinate functions, 764

coordinate representation of L8*, 820 coordinate transformation orientation preserving, 800 orientation reversing, 800 cosel,658 cosmologicalconstant, 921 Coulomb, 985 countabiy infinite set, 12 covariant degree, 731 covariant derivative, 897-908 and curvature, 903 Lie derivative, 902 covariant tensor, 731 criterion for irredncibility, 685 crystallography, 675 current density, 992 energymomentum, 998 curvature and gravity, 914 as relative acceleration, 913 matrix, 888 two-form, 885 curve coordinate, 774 differentiabie, 770 curvilinear coordinates, 892 cycle, 666

expansion Fourier, 204 general, 189 limit of seqnence, 161, 163 potential, 601 dense snhset, 460 density function, 833 density of states, 532 derivation, 840 derivation algebra, 840 derivation of an algebra, 42 derivation property tangent vector, 772 derivative covarianl, 897

derivative operator nnboundedness of, 455 Descartes, 738 detenrdnant, 93-101, 744 analytic definition of, 137

connection with trace, 102 definition, 93 derivative of, 102 expansion of, 95 exponential of trace, 103 of an operator, 101

products of matrices, 97 relation to trace, 101 diffeomorphism, 768 dUferentiablecurve, 770

tangent vector, 772 differentiable manifold, 763-769

INDEX

dimension of a, 764 differentiable map, 768

Frobenius method, 371 general properties, 352 integrating factor, 364 regular, 353

coordinate expression of, 768 differential of a constant map, 777 ofamap, 776 one-form, 786 real-valued maps, 778 differential equation analytic, 401 analytic properties, 401 associated Legendre, 343 Bessel,407

uniqueness theorem, 354 Wronskian, 356 symmetry group of, 941-950 differential form, 791

Maxwell'sequations, 793 differential forms Lorentz force law, 794 differential operator, 349, 863

differentiation operator, 73

confiuent hypergeometric, 420 constant coefficients, 376-379

diffusion equation, 528, 591, 621 time-dependent, 530

Euler, 411

diffusionoperator

first-order existence, 350

Green's function for, 632 dilation, 238 dilitation, 912

linear, 351 uormal form, 351

Peanoexistencetheorem, 351 uniqueness, 350 uniqueness theorem, 352 first-order linear irregular singular point, 402 regular singular point, 402 Fuchsian, 411 higher-order

numerical solution, 393 homogeneons, 349

hypergeometric, 407

dimension theorem, 35 Dirac, 850 biography, 167

Diracdeltafunction, 160 direct products, 661-{j63 direct sum, 109-111 definition, 110 directional derivative, 976 DUichlet, 177,738,888,891 biography, 614 Dirichlet boundary condition, 591 DUichlet BVP, 591, 613-{j19

in two dimensions, 638

definition, 413

Jacobi fUnctions, 418

discrete Fourier transform, 217-219

Kummer's solutions, 418

dispersion relation with one subtraction, 308 dispersion relations, 306-309

inhomogeneous, 349 linear, 349 superposition principle, 354

numerical solution, 384 Adam's method, 386 Euler's method, 386 Riemann, 412

secondorder linear hehavior at infinity, 410

canonical basis,404 characteristic exponents, 406 second-order linear adjoint, 365

distribution, 166 density, 166 derivative of, 169 Fourier transform of a, 219 limit of functions, 168 divergence, 987 oftensors, 903

domain,5 dual basis, 38, 729

of anoperator, 39

1011

1012

INDEX

space, 34

eigenfunction expansion technique, 6364/1 2D Laplacian, 638 eigenspace, 115 eigenvalue, 114-116 characteristic polynomial, I IS definition, 114 of angular momentum, 337 simple, 115 eigenvector, 114-116 angular momentum, 338-346 definition, 114 generalized, 468 Einstein, 799, 849, 889, 891, 899, 920, 921,996

Einstein tensor, 919 Einstein's equation, 918-921

Schwarzschild solution, 924 Einstein's summation convention. 728 electromagnetic field tensor, 796, 797

elementary column operation, 99 elementary rowoperation, 99 elliptic PDE, 589, 613-fj21 elsewhere, 838 empty set, 2 endomorphism, 33 energy function, 805 entirefunction, 233

equivalence class, 3 equivalence relation, 3 equivalent representations, 674 essential singularity, 273 essentially idempotent, 689 ~ -orthogonal matrices, 837 Euclid, 151,807 Euclidean metric, 892 Euclidean space, 754 Euler, 233, 390, 415, 423, 514,888,984 biography, 981 Euler angles, 89, 865 Euler kernel, 434 Euler operator, 980 Euler theorem, 865 Euler transform, 433 Euler-Lagrange equations, 981

Euler-Mascheroni constant, 310 evaluation function, 977 event, 837

evolution operator, 58, 626 exact form, 797 expectation value, 64 exponential function complex, 234 exponential map, 824 exterior a1gehra, 739--749 exterior calculus, 791--801

exterior derivative, 791 exterior product, 740

F-related vector fields, 780 faithful representation, 674 Fermi energy, 532 Feynman diagram, 602 Feymnan propagator, 636 field, 20 finite-difference operators, 70-73 finite-rank operator, 464 first integral, 992 first variation, 983 fiat manifold, 895, 896 fiat map, 747, 802 fiow,784 form factor, 216 formal adjoint, 560 formally self-adjoint, 560 forward differenceoperator, 70 Fourier, 614, 653, 982 biography, 198 Fourier series, 196-208 angnlar variable, 197 fundamental cell, 198 general variable, 200 group theory, 853 higher dimensions, 207 main theorem, 202 sawtooth, 201 square wave, 200 to Fourier transform, 208-210 Fourier transform, 208-220, 433 and derivatives, 216-217 and GF, 628-fj36 and quark model, 216

INDEX

Coulombpoteotial charge distribution, 215 point charge, 214 definition, 210 distribution, 219 Gaussian,212 higherdimensions, 213 Fourier-Bessel series,536 Fredhohu, 151 Fredhohu alternative, 496 Fredholmequatiou, 488, 600 Fredhohu, biography, 496 Friedmannmetric, 892 Friedmann,biography, 921 Frobenius,681, 850, 874 biography, 711 Frobenius method, 370-376 Fuchsian DE, 41()...413 definition, 411 function, 5 function algebra, 42 function of operator, 53, 125-128 derivative, 56 functional, 980 linear, 34 functional derivative, 977 functions of trig functions integrals of, 281 fundamental solution, 597-{;00 fundamentaltheoremof algebra, 251 future light cone, 837 G-invariance, 937 g-orthogonal,751 g-orthononruU,7~2

Galois, 96, 711, 842, 942 biography, 653 gamma function, 181,309-312 definition, 309 gaugeinvariance,797 Gauss, 96, 183,233,422,463,478,614, 738,797,888,891,981 biography, 415 Gay-Lussac, 528 Gegenbauerfunctions,419 Gegenbauerpolynomials, 185 general linear group, 655

1013

representation, 856-859 general relativity,918-932 generalized eigenvector, 468 generalized Fourier coefficients, 150 generalizedfunctions, 165-169 generalized Green's identity, 561, 574, 597 generating function, 189 geodesic, 905-908 relative acceleration, 913 geodesic deviation,913-918 equation of, 913 geodesic equation massive particles, 926, 927 massless particles, 926, 928 geometric multiplicity, 115 Gibbs, 463, 807 Gibbs phenomenon, 205-207 GL(n, lll.) as a Lie group, 816 GL(V) as a Lie group, 816 representation of, 856 GOde1,799 Gordan, 996 biography,704 gradient for Hilbert spaces, 976 gradient operator, 901 Gram, 28 Gram-Schmidt process, 27-29 Gram-Schmidtprocess, 172 graph, 5 Grassmann,743, 996 Grassmann product, 740 gravitatioualred-shift, 930 gravity and curvature,914 Green'sfunction, 289 adjoint, 566 advanced,634 andFourier transform, 628 construction, 569 diffusion operator, 632 Dirichlet BC circle, 618 Dirichlet BVP,613 eigenfunctionexpansion, 577-579

1014

INDEX

for SOLDO, 565-577 for the Laplacian, 595-596 formal considerations, 557 Helmholtz operator, 630 in 20, 643 in one dimension, 553 indefinite, 554 inhomogeneous Bes, 574 iutegral equations, 600 Laplacian, 629 NeumannBVP,620 exterior, 621 interior, 620 physical interpretation, 577 properties, 567 retarded, 634 second order DO, 562 synnnetry, 567 wave equation, 633 Green's function for dhlx, 554 for d 2/ dx 2, 555 Green's identity, 567, 596, 623, 627 Green, biography, 561 group, 652-656 abelian, 654 affine, 876 algebra, 687 symmetric gronp, 718 automorphism, 655 center of a, 658 commutative, 654 commutator, 657 homomorphism, 655 isomorphism, 655 left action, 663 multiplicatiou, 652 multiplication table, 656 of affine motioos, 817 order of, 652 orthogonal, 656 realization, 664

representation definition, 674

irreducible, projection operator, 698 matrix,675

particles and fields, 699 teusor product, 699 right action, 663 rigid rotations, 657 special orthogonal, 657 special unitary, 657 symplectic, 657, 749 unitary, 656 group action, 663-664, 817-820 effective, 663 infinitesimal generator, 827 transitive, 663, 818 guidedwaves TE,533 TEM,533 TM,533 Haar measure, 832 Halley, 422 Harrrilton, 178,490,996 biography, 806 Hamiltonian system, 805 Harrriltonian vector field, 805 Hankel function, 425 asymptotic expansion of, 316 Hankel transform, 434 harmonic functions, 236 heat equation, 327, 591, 621 synnnetry group, 957-961 heat transfer, 528 steady-state, 528 time-dependent, 530 heat-conducting plate, 528 Hegel,738 Heisenberg, 64, 167 helicity, 876 Helmholtz, 178, 588, 849 Helmholtz equation, 542 Helmholtz operator Green's function for, 630 Hermite, 183,798 biography, 63 Hennitepolynomials, 176, 180, 517 hermitian conjugate, 61 hermitian kernel, 497-501 hermitian operator, 63-67 llilbert, 11,28, 199,463,704,799,848, 920,996

INDEX

biography, lSI Hilbert space,145-157 basis, ISO definition, 148 differential of functions, 976 functions on derivative of, 974 Hilbert transforms, 308 Hilbert-Schmidt operator, 465 Hilbert-Schmidt theorem, 498 Hilbert-Schmidt kernel, 466, 494 Hilbert-Schmidt operator, 847 Hodge star operator, 756-758, 795 HOlder,463 homogeneous SOLDE

exact, 364 second solution, 358 homographic transformations, 239 homomorphism algebra, 43 kernel,658 Lie group, 816 synnnetric, 655 trivial, 655 hyperbolic PDE, 589, 626-628 hypergeometric DE, 407 hypergeometric function, 413-419 contiguous, 417 Euler formula, 436 integral representatioo of, 434 hypergeometric series,414 hypersurface, 583 ideal,44 idempoteot, 688 esseotially, 689 primitive, 689 identity additive, 20 multiplicative, 20 ideotity map,S identity operator, 50 identity representation. 674 iguorable coordinate, 593 imageof a subset, 5 implicit function theorem, 350 iodicial polynomial, 406

inducedrepresentations, 871 induction principle, 12 inductive definition, 14 inequality Bessel,I49 Darboux, 245 Parseval,I49 Schwarz, 29 triangle, 31 infinitesimal actioo, 826-831 adjoint, 827 infinitesimal generator, 827, 830 inhomogeneous B'Cs, 574-577 inhomogeneous SOLDE general solution, 360 initial value problem, 559, 583 numerical solution, 383 injectivemap, 6 inner automorphism, 825 innerproduct, 23-32 braandket notation, 25 definition of, 24 norm and, 31 positive definite, 25 pseudo-Riemannian. 25 Riemannian, 25 sesqnilinear, 25 innerproduct space, 25 integrability condition, 888 integral curve, 783 integral equation characteristic value, 489 Fredhohu,494-504 Green's functions, 600 kernel of, 488 of the first kind, 488 of the second kind, 488 Volterra, 488 Volterra, of secondkind solution, 490 integral operator, 452 integral transforms, 433 Bessel function, 434 integration on manifolds, 800-801 integration operator, 73 interior product, 793 intersection, 2

1015

1016

INDEX

intrinsic spin, 999 invariant, 937 invariant map, 937

invariant operator matrix representation, 112 invariant subspace, 112-114, 676, 677 definition, 112 inverseimage, 5 inverse mapping theorem, 777

inverse of a map, 6 inverse of a matrix, 97-100 inversion, 238, 912 irredncibility criterion, 685 irreducible representation, 677

i-th row functions, 696 norm of functions, 696

irreducible tensor operators, 705-707 isometric map, 908 isometry, 90&-912 timetranslation, 923 isomorphism, 36 algebra, 43 Lie algebra, 822 Lie group, 816 Jacobi, 183,416,490,615, 663, 702, 704, 738,807,888 biography, 177 Jacobi functions, 418 Jacobi identity, 782, 790, 826 Jacobi polynontials, 176, 184,418

special cases, 176 Jacobian matrix, 777 Jordan arc, 242 Jordan canonical form, 485 Jordan's lemma, 275

biography, 842 Killing equation, 909 Killing form, 841, 845 Killing form of glen, R), 844 Killing vector field, 90&-912, 923, 926, 930 conformal, 911 Kirchhoff, 588 Klein, 388, 743, 798,849,891,920,942, 996 Klein-Gordon equation, 328 Korteweg-de Vries equation, 971 Kovalevskaya, 463 biography, 587 Krarners-Kronig relation, 309 Kronecker, 11,96,388 hiography, 738 Kronecker delta, 26 . Kummer, 30, 704, 738, 842, 891. KUlla, biography, 390 Lagrange, 96, 178, 182, 198, 199,415, 423, 528, 704 biography, 984 Lagrange identity, 366, 434, 525, 560 Lagrange multiplier, 991 Lagrange theorem, 670 Lagrangian, 804, 980 null, 987 Laguerrepolynontials, 176, 180-181 Laplace, 28, 199,614,806,920,982,985 biography, 527 Laplace llansfonn, 433 Laplace's equation, 327

Cartesian coordinates, 526 cylindrical coordinates, 535

Kan~ 738 Kelvin, 561 Kelvin equation, 538 Kepler problem, 1001 kemel, 35 degenerate, 501 integral operator, 452 integral transforms, 433 separable, 501 Killing, 743, 942

spherical coordinates, 541 Laplacian Green's function for, 595, 629 separation of angular part, 331 Laurent series, 252-262, 606 construction, 254 uniqueness, 258 Laurent, biography, 271 Lavoisier, 95, 985 left coset, 658

INDEX

left ideal, 44, 689 minimal, 689 left translation, 820 as action, 827 left-invariant l-form, 820 left-invariant vectors, 820 Legendre, 177, 199,490,614,888 biography, 182 Legendre eqnation, 517 Legendre functions, 419 Legendre polynontials, 155, 182-184, 340,359 and delta function, 188 and Laplacian, 188 Legendre transformation, 804 Leibniz, 95, 738 Leibniz rule, 15 length of a vector, 30-32 Levi-Civita, 891 biography, 898 Levi-Civita tensor, 745, 869 Lie,711,743,798,842 biography, 941 Lie algebra, 833-845 abelian, 834 adjoint map, 825 Cartan metric tensor, 843 Cartan theorem, 844 center, 834 commutative, 834 compact, 843

decomposition, 844 derivation, 840 ideal, 834 Killing form of a, 841 of a Lie gronp, 820-826 of nnitary gronp, 824 of vector fields, 782 representation, 859--876 definition, 845 semisimple, 844 simple, 844 structure constants, 834 Lie algebra of SL(V), 823 Lie bracket, 782 Lie derivative, 788

covariant derivative, 902 of a I-form, 790 of vectors, 789 Lie derivative of p-forms, 793 Lie gronp, 815 compact characters, 853 matrixrepresentation, 852 representation, 852 unitary representation, 846 Weyloperator, 847 homomorphism, 816 integration, 832-833 density function, 833 local,817 representation, 845 Lie mnltiplication, 833 Lie snbalgebra, 834 Lie's first theorem, 829 Lie's secondtheorem, 826 Lie's third theorem, 826 ligbtcone, 837 linearcombination, 21 linear functional, 34, 455 linearindependence, 21 linear operator, 33 nnll space of a, 35 linear POE, 584 linear transformation, 32-40 bonnded,454 definition, 33 pullback of a, 39 Liouville, 512, 654 biography, 514 Liouville substitution, 513, 520 Liouville's theorem, 807 Lipschitz condition, 351 little algebra, 871 little group, 871, 873 local diffeomorphism, 768 local group of transformations, 817 local Lie group, 817 local operator, 452 logarithntic function, 295 Lorentz, 799 Lorentz algebra, 864

1017

1018

INDEX

Lorentz forcelaw,794 Lorentz group, 657, 837 Lorentz metric, 892 Lorentztransformations, 837 loweringindices, 750 lowering operator, 335

Maclaurin series, 254 manifold, 764 coordinate functions, 764

ffat, 895 orientable, 800 product, 767

pseudo-EUemanmrran,887 symplectic, 801 map, 4-7 bijective,6 codomain,S dlfferentiable,768 domain, 5 equality of, 5 functions and, 5 graph of a, 5 identity, 5

imageof a subset, 5 injective, 6 inverse of a, 6 one-to-one, 6 onto, 6 range of a, 5

surjective, 6 target space, 5 matbematical induction, 12 matrix, 82-86 antisymmetric, 87 basis transformation, 91

complexconjugate of, 87 diagooal, 88 hermitian, 87 hermitian conjugate of, 87 inverseof a, 98 irreducible, 112

operations on a, 87 orthogonal, 87 rank of a, 86 reducible, 112, 113 representation

orthonormal basis, 89 row-echelon,99 symmetric, 87 transpose of a, 87 triangular, 99 unitary, 87 matrix algebra, 42 maximally symmetric spaces, 910 Maxwell's equations, 796 Melliu transform, 434 Mendelssobn,615,738 meromorphic functions, 293-294 metbod of images, 616 sphere, 617 methodof steepest descent, 521 metric space, 7-10 convergence, 9 definition, 8 minimal ideal, 44, 856 Minkowski, 920 Minkowski metric, 892 Minkowski space, 754 Mittag-Leffler, 388, 463, 588 Mittag-Leffler expansion, 294 modified Bessel functions, 425 moment of inertia matrix, 88 Monge, 95, 198 Morera's theorem, 252 multilinear mapping, 729-735 tensor-valued, 735 multiplicative identity, 20 multivalued functions, 295-302 n-equivalent functions, 945 a-sphere, 764 n-tbjet space, 945 Napoleon, 198, 527 natural isomorphism, 733, 756 natural pairing, 731 neighborhood, 459 Neumann, 177, 702 biography, 620 Neumann Be, 591 Neumann Bvp, 591, 619-621 Neumann function, 424 Nemnarm series, 493, 601, 602 Newton, 330,415, 527,798, 806,982

INDEX

Newtonian gravity,916-918 Noether,704 biography, 996 Noether's theorem, 995 non-local potential, 631 norm of a vector, 30 operator, 454 product of operators, 456 normal operator, 117 nornuU snbgronp,659 normal vectors, 26 normed linear space,31 normed vectorspace compact subsetof, 461 noll divergence,993 noll Lagrangian, 987 noll space, 35 noll vector, 751, 837 numerical integration, 74-76 Simpson's 1/3 rule, 75 Simpson's 3/8 rule, 75 trapezoidal rule, 75 Ohm,614 Olbers,422 one-form, 786 one-parameter group, 784 open ball, 459 open subset, 459 operations on matrices, 87-89 operator, 33 adjoint, 61 existence of, 457 angular momentrun, 331 annihilation,374 bouuded,454 closed,508 bounded, 508 compact Hermitian spectral theorem, 476 compact normal spectral theorem, 477 compact resolvent, 508 creation, 374 diagonalizable,116 differentiation, 73

domain of, 507 eqnality,49 evolution, 58 expectation valueof, 64 extension of, 508 finite rank, 464 formally self-adjoint,597 functions of an, 53 hermitian, 63, 508 eigenvalue, 117 hermitian conjugate of, 61 Hilbert-Schmidt, 465, 511 integration, 73 inverse of an, 50 kernel of an, 35 local,452 negativepowers of, 52 norm of an, 454 normal,117 diagona1izable, 119 eigenspace of, 118 noll space of an, 35 numerical analysis, 70 polarization ideotity,50 positive,65 projection, 67 projection orthogonal,68 pollback of an, 39 raising andlowering, 335 regular point, 457 representation of an, 83 right-shift,453 scalar, 706 self-adjoint,508 spectrnm,457, 457-458 spectrnmof an, 115 square rootof, 127 Stnrm-Liouville,510 trace of an, 102 unbounded,507 mlitary, 66, 127 eigenvalue, 117 operator algebra, 49-76 operator polynomials, 51-53 operator spectrum, 457 o(p, n - pl, 837

1019

1020

INDEX

optical theorem, 308 orbh, 663,676,818 orbital angular momentum, 999 ordered pairs, 2 orieiltable manifolds, 800 orientation, 745,746 positive, 746 orientedbasis, 745 orthochronous, 837 orthogonal groop, 656, 824 Lie algebra of, 824 orthogonal polynomials, 153 classical, 173 classification, 176 differential equation, 174 generating functions, 189 recurrence relations, 176 expansion, 186 least sqnare fit, 155 orthogonal transformation, 97 orthogonal vectors, 26 orthogonality, 26-27 orthonormal basis, 26 orthonormal frames, 887

periodic Be, 515 permutation even, 668 odd,668 parity of, 667 permutation group, 664 permntation symbol, 93 permutation tensor, 754 perturbation theory, 603-{jlO, 697 degenerate, 609 first-order, 608 nondegenerate, 607 second-order, 608 Peter-Weyl theorem, 853 Peter-Weyl theorem Fourier series, 853 photoncapture

crosssection, 929 Planck,388,463,920 Poincare,63,478,497,620,744,920 biograpby, 797 Poincare algebra, 840, 844 representation, 868-876 Poincare group, 657, 817, 839, 872

Poincare lemma, 797 p-forms, 741 pairing natural, 731 parabolic PDE, 589, 621-{j26 parallel transport, 906 parallelogram law, 31 parity,667 Hermite polynomials, 194 Legendre polynomials, 194 Parseval inequality, 149, 851 Parseval's relation, 223 partial DE, 584-592 characteristic system of, 939 order of, 584 principal part, 584 particle in a box, 531 particle in a cylindrical can, 539 particle in a bard spbere, 543 partition,4,669 past light cone, 838 Pauli spin matrices, 88, 835, 840 Peano, 799

converseof, 797 Poisson, 178,512,528,614,653 Poisson bracket, 808 Poisson's equation, 327, 917 polar decomposition, 129-130 pole, 273 polynomials orthogonal, 153

positiveoperator, 65 positive orientation, 746 potential non-local, 631 separable, 631

power series, 252 differentiation of, 253 integration of, 253

uniform convergence, 253 p(p, n - p), 837 preimage,5

primitive idempotent, 689 principal value, 285 product manifold, 767

INDEX

projectable symmetry, 944 projection operator, 67-70, 474, 604 completenessrelation, 68 orthogonal, 68

projective group density function, 833 one-dimensional, 820 projective space, 4 prolongation, 945-950 of a function, 946 of groups, 948 of vector fields, 949 propagator, 626 Feynman, 636 proper orthochronous, 837 proper subset, 2 Priifer substitution, 518 Puiseux, biography, 296 pullback,735,786,791,801

linear transformation, 39 pullback map, 791 pullback of p-forms, 741

quantization harmonic oscillator algebraic, 375 analytic, 374 hydrogen atom, 422 quarks,70I,703,873 quaternions, 44 quotient group, 660

quotient set, 4 raising indices, 750 raising operator, 335 range of a map, 5 rank of a matrix, 86 rational function, 274 and trig functions integrals of, 279 integration of, 276

rational numbers dense subsetof reals,460 real coordinate space, 21 real vector space, 20, 130-137 realization, 664 reciprocal lattice vectors, 208

recurrence relations, 153

redshift, 930 reducedmatrixelements, 707 reduciblerepresentation, 677 regular graphs, 713 regular point, 233 regular representation, 686 regular singular point second-order linear DE, 405 relation, 3 relativeacceleration, 913 relativisticelectromagnetism, 792 representation abelian groop, 681 adjoint, 679, 704 carrier space, 674 cbaracter, 683 compact Lie group, 852 complex conjugate, 679 dimension of, 674 direct sum, 679 equivalent, 674 faithful, 674 geuerallinear group, 856-859 group, 674 identity, 674 irreducible, 677 Kronecker product, 700 Lie group, 845 operators, 83 reducible, 677 regular, 686 symmetric group, 707-723 tensor prodoct, 700 antisynunetrized, 701 character, 700 synunetrized, 701 trivial, 679 unitary, 678 compact Lie group, 846 representation of g[(n,lR), 860 s[(n, It), 861 so(3),864 so(3, I), 867 su(n),861 u(n),861 residue, 270-273

1021

1022

INDEX

definite integrals, 275 definition, 271

residue theorem, 271 resolution of identity, 482, 689, 721 resolvent, 481J--485 compact, 508

Laurent expansion, 481 resolventset. 457 openness of, 460 resonant cavities, 533 Riccati equation, 968 Ricci, 891, 898 Ricci tensor, 917, 919, 920 Riemann, 30, 199,296, 704, 798,849, 891,981 biography, 888 Riemanncurvature tensor, 889, 904 Riemann normal coordinates, 915 Riemann surfaces, 296-302 Riemannian manifold, 887-896 right coset, 658 right ideal, 44 right translation, 820 right-invariant l-form, 820 right-invariant vectors, 820 right-shift operator, 453 eigenvalues of, 458 rigid rotations, 657 Rodriguez formula, 174 Rosetta stone, 198 rotation algebra, 864 , rotation gronp, 675, 862 character, 865 rotation matrix, 865 Runge, biography, 388 Runge-Kntta method, 387~391 Rnssell, 11, 799 saddle point approximation, 313 scalar operator, 706, 707

scale transformations, 819 Schelliog,738 Schmidt, biography, 28 Schopeohaner,738 Schriidinger, 64, 807 SchrOdinger equation, 675 Schriidinger equation, 58, 328, 410, 420, 531,543,631

classical limit, 383 one dimensional, 381 Schur, 711, 850, 874 biography, 681 Schur's lemma, 680, 681, 707,845,861 Schwarz, 463, 739 biography, 30 Schwarz inequality, 29, 149, 152 Schwarz retlection principle, 305-306 Schwarzschild geodesic, 925-932 Schwarzschild metric, 892 Schwarzschild radius, 925 Schwarzschild, biography, 919

secondorder linear DE Sttum-Liouvi11esystems, 513 second order POE, 589--592 hyperbolic, 589 second-order PDE classification, 589 elliptic, 589 parabolic, 589 ultrahyperbolic, 589 selection rules, 701

self-adjoint operator, 63 self-adjoint SOLDOs, 564-565

semisimple Lie algebra, 844 separable kernel, 501 separable potential, 631 separated boundary conditions,510

separation of variables, 328 Cartesian, 526-534 cylindrical, 535-540 spherical, 540-544 separation theorems, 361-363 sequence, 9 Canchy,1O complete orthonormal, 149

series Fourier-Bessel, 536 ~eornann,601,602

sesquilinear inner product, 25 set, 1-3 Cantor, 12

complement of, 2 countably infinite, 12

element of, 1 empty, 2

INDEX

intersection, 2 partition of a, 4

uncountable, 12 union, 2 universal, 2 sharp map, 747, 802 shifting operator, 71 signature of 9,752 similarity transformation, 91-92 orthonormal basis, 91 simple arc, 242 simple character, 685 simple Lie algehra, 844 simultaneous diagonalizahility, 123 singleton, 2 singular point, 233 singularity, 233 isolated, 270 classification, 273 rational function, 274 skew-symmetric, 738 SL(V) as a Lie group, 816 smooth arc, 242 space

Banach, 148

cornplexcoorilinate,21 dual, 34 inner product, 25 metric, 8 real coordinate, 21

square-integrahlefunctions, lSI spacelike vector, 837 spacetime spherically symmetric, 923 static, 923 stationary, 923 span, 22 special linear group, 656 special orthogonal group, 657, 824 Lie algebra of, 824 special relativity,752,837,868,872,986 special unitary group, 657 Lie algebra of, 824 spectral decompusition,117-125 orthogunal operator, 135 real case, 131 spectral theorem

1023

compact hermitian, 476 compact normal, 477 spectrum closure of, 461 spherical Bessel functions, 428, 542 expansion of planewave, 543 spherical coordinates in TIl dimensions, 593-595 spherical harmonics, 338-346 addition theorem, 345, 867 definition, 341 expansion in terms of, 344 expansion uf plane wave, 544, 646 first few, 343 Spinoza, 738 square-integrable functions, lSI stabilizer, 663, 818 staticspacetime, 923 stationary spacerime,923 steepest descent method, 312-315 step function, 163,288 stereographic projection n-sphere, 769 two-sphere, 766 Stirling approximation, 316 Stone-Weierstrass theorem, 152 generalized, 196 stressenergytensor, 920 structure constants, 826, 900, 905 structure equations, 888 Sturm, biography, 512 Sturm-Liouville operator, 510 problem, 174,622 system, 510, 637 completeness, 524 eigensolutions, 511 eigenvalues, 513 large argrunent, 521 largeeigenvalues, 518 regular, 510 singular, 516 subalgebra, 44 subgroup, 656-663 conjugate, 657 generated by a subset, 657 normal, 659

1024

INDEX

trivial, 656 submanifold, 767 open,767 subset, 2 bouoded,459 proper, 2 sobspace, 22 invariant, 112 surjective map. 6 Sylvester's theorem, 753 symmetric algebra, 737 symmetric bilinear form, 749 classification, 751 inner product, 750 nondegenerate.vsn symmetric group, 655, 664-
graphical construction, 714 cycle, 666 representation, 707-723 analytic construction, 707 antisynnnetric, 679 graphical cnnstruction, 710 products, 721 Young operators, 717 symmetric bomomorphism, 655 symmetric prodnct, 737 symmetrizer, 736 symmetry group, 944 defining eqnations, 957 of a DE, 941 of a snbset, 937 projectable, 944 transform of a function, 943 variational, 988 symplectic algebra, 836 symplectic cbarts, 802 symplectic form, 746, 801 rank of, 747 symplectic geometry, 801-808 conservation of energy, 805 symplectic group, 657, 749, 835 symplectic manifold, 801 symplectic map, 746, 802 symplectic matrix, 749 symplectic structnre, 801 symplectic transformation, 746

symplectic vector space, 746-749 canonical basis, 747 Hamiltonian dynamics, 748 tangent buodle, 780 tangent space, 773 tangent vector, 772 tangents to a curve components, 778 target space, 5 Taylor series, 252-262 construction, 254

tensor, 731 components of, 732 contravariant, 731

contravariant-symmetric,736 covariant, 731 covariant-symmetric, 736 Levi-Civita, 745 skew-symmetric, 738 symmetric, 736 transformation law, 733 types of, 731 tensoralgebra, 731 tensorbundle, 786 tensor field, 785-790 crucial property of, 786, 787 tensor product, 730, 731 of vectorspaces, 41 tensors symmetric product, 737 theta function, 288 timelike vector, 837 tntal derivative, 954 total divergence, 987 trace, 101-103 and determinant, 102 definition, 101 log of determinant, 103 relation to determinant, 101 transformation group, 655 translation, 819 translation operator, 138 transpose of a matrix, 87 transposition, 667 traveling waves, 533 triangle inequality, 31

INDEX

trivial homomorphism, 655 trivialrepresentation, 679

trivial suhgroup,656 twin paradox as a variatioualprohlem,986 uncertainty principle, 78 uncountable set, 12 union, 2 unitary group, 656 Lie algebra of, 824 unitaryoperator, 63-
Vaudermonde, biography, 95 variation of constants, 360 variational derivative, 977 variationalproblem,980 twin paradox, 986 variationalsymmetrygroup, 988 vector field, 780 flowof a, 784 integral curve of, 783 left-invariaut, 820 Lie algebra of, 782 vector space, 19-45 automorphism, 36 basis components in a, 22 basisof a, 22 complete, 146 complex, 20 definition, 19 dual, 34 endomorphism of a, 33 finite-dimension criterion for, 462 finite-dimensional, 22 isomorphism. 36 linear operator on a, 33 normed,31 operator on a, 33 oriented,746 real, 20 symplectic, 746 vectors, 19 norm of, 30

1025

uormal,26 orthogoual, 26 Volterraequatious,488 Volterra, hiography, 490 volumeelement, 746 vou Dyck, 390 vou Hmuholdt, 177,614,738 von Neumann, 874

hiography,478 wave equatiou,328, 532 symmetrygroup, 961-1J63 wave guide, 533 cylindrical,537 rectangular,534 wedge product, 740 VVeierstrass, 11,30,296,388,588,739, 841 biography, 462 VVeyl,744, 842, 942, 996 biography, 849 VVeyl basis, 835, 844 Weyl operator,847 VViguer, 168,943 biography, 874 VViguer formula, 865 VViguer-Eckart theorem, 707 Wigner-Seitzcell, 208 \VKB method, 380-383 connectionformulas, 381 VVordsworth, 806 Wronski, biography, 357 VVrouskiau,355-363,511 comparison theorem, 362 separationtheorem, 361 Young, 850 Youngframe, 712, 718 negativeapplication, 715 positive application, 715 regular application, 714 Young operator,856 Young pattern,712 Youngtableanx,713, 718, 857 Yukawapotential, 214